paperclip/doc/connections/CONNECTOR-PLAYBOOK.md

103 KiB

Connection Authoring Runbook

Audience: agents and engineers researching, implementing, testing, reviewing, and shipping Paperclip app connections.

Status: canonical end-to-end authoring guide for Apps v2 catalog connections.

This runbook is the repeatable, agent-executable procedure for adding a vendor to the Apps catalog as data, not as a plugin. It follows the accepted connections framework in PAP-13211, the first-30 rollout matrix in PAP-2432, and the production validation scope in PAP-12373.

Use it when Paperclip acts on an external system through a governed connection: a stored credential, a capability catalog, access profiles and policy rules, and audit. Inbound integrations, such as an external client acting on Paperclip, use gateway or webhook guidance instead.

A catalog entry is a convenience layer, not a prerequisite. Since PAP-17087, an operator can connect any standards-compliant remote HTTP MCP server from Connect your own MCP server or Paste a config with no Paperclip code change at all — including servers that need browser sign-in. Those two routes are the documented baseline; see Connecting any remote MCP server.

Write a catalog entry when Paperclip should promote a vendor: branding, tailored fields, field validation, scoped defaults, and support copy. A definition adds those conveniences and nothing else. It must not create a second connection, change ownership, or be necessary for health, catalog, or governance — a curated route and the generic route converge on the same connection and review pipeline.

Every connector built with this playbook is a plane P2 connection — a resource credential governed by the Paperclip instance, never a sign-in authenticator. The default durable authority is the instance vault. Reviewed remote MCP methods may opt in to Vercel Connect, where durable provider credentials remain in the operator's Vercel account and Paperclip resolves short-lived tokens at invocation time. Before writing a connector, read Identity vs. connections for the P1/P2/P3 boundary and the D7 standing rule.

AI provider credentials use the same vault, applications, grants, installations, and delegation model with connectionPurpose: ai and transport: runtime_auth. They authenticate provider execution and never enter MCP discovery or tool/channel execution. Extend the provider's existing catalog entry with typed AI methods; reuse the existing login controllers. See AI Connections for compatibility, personal defaults, resolver isolation, and legacy adoption.

Contents

Output

A complete connector proposal produces:

  • A catalog manifest entry with user-facing app metadata.
  • Transport and auth configuration.
  • Credential secret refs into company_secrets; never raw env values.
  • Action catalog metadata with risk classes, schemas, resource filters, and quarantine defaults.
  • Default profile and policy behavior for read, write, and destructive actions.
  • A smoke checklist aligned to PAP-12373: connect, discover catalog, allowed read call, correctly governed write call, denied/quarantined call when the method declares one, revoke, and audit evidence.

Use This Document As The Checklist

An agent implementing a connection should be able to begin with only a provider name and this document. Work in order. Do not jump from finding an MCP URL to adding a store card; the research, credential, risk, branding, deterministic test, live proof, and PR steps are all part of the feature.

The shortest valid implementation usually changes these files:

scripts/ingest-app-definitions.mjs                # human-authored definition source
packages/shared/src/app-definitions/<slug>.json  # generated definition
packages/shared/src/app-definitions.generated.ts # generated registry
ui/public/brands/apps/<slug>.svg                  # official, sanitized mark
ui/public/brands/apps/manifest.json               # branding provenance
packages/shared/src/app-definitions.test.ts       # manifest/provider assertions

Add server or UI code only when the provider cannot be represented by the existing contract. Prefer extending one generic capability with fixtures over adding a provider-name branch. Provider branches are justified for behavior that cannot be inferred safely, such as a provider's reviewed risk exceptions, permanently blocked actions, or a protocol-required managed argument.

The rest of this document has three levels:

  1. Mental model and support matrix — choose the right connection type.
  2. Golden-path tutorial — research, implement, test, prove, and submit.
  3. Reference and appendices — field semantics and worked examples.

Mental Model: Five Independent Axes

Do not describe a connection as merely "an OAuth connection" or "an MCP connection." OAuth is authentication. MCP is transport. A complete method chooses all five axes below.

Axis Current values Question
Transport mcp_remote, local_stdio, rest_api How does Paperclip reach actions?
Authentication oauth, api_key, none How does the provider authorize requests?
OAuth client ownership dcr, customer, platform_shared, platform_provisioned Who supplies and controls the OAuth client registration?
Credential source paperclip_vault, reviewed vercel_connect Where does durable provider credential material live?
Grant identity organization, user, agent Does the credential act for the company, one person, or one dedicated agent?

These axes produce combinations such as:

  • Remote MCP + DCR OAuth + Paperclip vault + organization identity: Jira.
  • Remote MCP + customer OAuth app + Paperclip vault: Asana.
  • Remote MCP + DCR or customer OAuth app: Notion and PostHog.
  • Remote MCP + API key in an HTTP header: Mem0 and PagerDuty.
  • Remote MCP + secret-bearing provider-generated URL: Zapier.
  • Remote MCP + no auth + required tenant field: Shopify.
  • Remote MCP + Paperclip-managed OAuth client + per-user grant: Google Workspace MCP previews.
  • Remote MCP + Paperclip-managed OAuth client + personal or dedicated-agent grant: GitHub. See GitHub managed connection.
  • Local stdio MCP + approved command template: the Google Sheets robot flow and development fixtures.
  • REST API parent + provider-specific child-session bridge: Composio. This is a specialized implementation, not a generic REST catalog recipe.

Transport support and boundaries

Transport Manifest-only? Runtime status Authoring rule
mcp_remote Yes First-class discovery, health, catalog, gateway, test, OAuth, and credential projection. Default for official hosted MCP servers.
local_stdio Only with an approved template First-class only through registered templates and a trusted runtime host. Disabled in authenticated/public deployments without that host. Never put an arbitrary command in an AppDefinition. Register and test a template.
rest_api No, not generally Not exposed through the connected MCP gateway. Composio is a provider-specific parent that creates MCP-capable children. Do not add a generic REST/API card until an execution adapter or wrapper exists.

api_key in a method means an authentication mode; it does not mean the transport is a REST API. Most current API-key catalog entries authenticate a remote MCP server.

Anthropic accounts use the runtime_auth AI connection methods. Its obsolete api-key REST tool method is no longer offered. Existing unsupported REST tool connections fail health and catalog checks with HTTP 422 and tool_connection_transport_unsupported; they never use local stdio templates or report a successful MCP probe. Add the provider through its supported account flow, then remove the obsolete connection. This does not transfer credentials or grants automatically.

For mcp_remote, header credentials and secret-bearing generated URLs have the complete generic runtime path. The schema also names query, body_json, and env key placements for specialized transports, but accepting a value in the schema is not proof that the remote MCP gateway projects it. Do not ship one of those placements without tracing the invocation path and adding an end-to-end fixture. env belongs primarily to approved local stdio templates.

Authentication support matrix

Pattern Definition shape What the user sees What Paperclip stores
Automatic OAuth auth: "oauth", ownershipModes: ["dcr"] Browser sign-in DCR/CIMD client binding plus token secret refs.
Automatic OAuth with own-app escape hatch ownershipModes: ["customer", "dcr"] Recommended browser sign-in; own client under Advanced Same as automatic, or supplied client ID plus encrypted client secret.
Customer OAuth only ownershipModes: ["customer"] Required client ID and optional/required client secret, then browser sign-in Client ID in redacted config; client secret and provider tokens as secret refs.
Paperclip-managed OAuth oauthStrategy: "paperclip_cloud_connector", connectorProfile, platform_shared Browser sign-in through Paperclip Cloud Provider tokens still land in the instance vault on a user grant. Cloud handles the fixed provider callback but does not persist plaintext credentials; Paperclip ID remains identity-only.
API key/PAT auth: "api_key", credentialFields, keyPlacement Write-only credential field Encrypted secret version plus placement-only refs.
Generated URL auth: "none", no fixed URL/default template Paste provider-generated MCP URL Public URL shape in config; full secret-bearing URL in the vault.
No auth auth: "none", fixed serverUrl or validated serverUrlTemplate Zero fields or only required tenant/resource fields No provider credential.

OAuth client resolution order

For standard OAuth methods, Paperclip resolves a client in this order:

  1. Deployment-preconfigured provider client.
  2. Client ID Metadata Document (CIMD), when advertised and the instance has a public HTTPS URL.
  3. Dynamic client registration (DCR/RFC 7591), when advertised and allowed by the curated method.
  4. A customer-created client ID and secret supplied through setup.

For a curated method, ownershipModes is an allowlist. A method containing only customer must not silently fall through to generic DCR. A method containing dcr permits the automatic CIMD/DCR tiers. Deployment-preconfigured credentials still take precedence when present.

Credential source is not OAuth client ownership

ownershipModes says who owns the OAuth client registration. It does not say where provider access tokens live. By default, access tokens, refresh tokens, client secrets, and API keys live in the Paperclip instance vault.

credentialSource: "vercel_connect" is a separately reviewed exception for specific methods. Such a connection stores a Vercel connector reference and no Paperclip provider secret refs. Never make a method accept both sources in the same connection, and never infer Vercel eligibility from a provider name.

Architecture From Setup To Agent Call

flowchart LR
    D["AppDefinition"] --> S["Shared validation and gallery"]
    S --> W["Connection setup flow"]
    W --> C["tool_connections draft"]
    W --> V["Instance encrypted vault"]
    V --> R["Credential secret refs"]
    C --> H["Health and catalog discovery"]
    R --> H
    H --> A["Risk-classified tool catalog"]
    A --> P["Profiles, bindings, policies"]
    P --> G["Run-scoped tool gateway"]
    G --> X["Provider MCP server"]
    X --> G
    G --> U["Redacted result and audit"]

The agent never receives a durable provider credential. A run receives a Paperclip gateway capability. At invocation time the gateway rechecks company, connection, grant, catalog, profile, policy, and run state; resolves the needed secret version; projects only the reviewed headers/arguments; calls the provider; and writes redacted audit evidence.

Secret Storage And Lifecycle

Connection credentials use the same secrets architecture as agent, project, and routine environment bindings. They are not files in a provider-named folder.

The durable pieces are:

  • company_secrets: secret identity, company/user scope, provider metadata, and ownership.
  • company_secret_versions: encrypted or externally referenced version material.
  • company_secret_bindings: which connection/grant/config path may resolve a secret.
  • secret_access_events: audited resolution activity.
  • tool_connections.credentialSecretRefs and connection_grants.credentialSecretRefs: value-free pointers and config paths.
  • tool_connections.credentialRefs: value-free projection shape such as the HTTP header name and prefix.

On the default local provider, values are encrypted with the instance master key under ~/.paperclip/instances/<instance>/secrets/master.key. A usable backup requires both the database and this key. Hosted provider-vault behavior is configured under Company Settings; the connection contract remains refs, not raw values.

Credential handling by pattern:

  • API keys are created through secretService.create, then only their refs are attached to the connection or selected grant.
  • OAuth access and refresh tokens use oauth.access_token and oauth.refresh_token refs. Refresh rotates versions under a lease so two servers do not replay a rotating refresh token.
  • Customer OAuth client secrets use an encrypted oauth.client_secret ref. Client IDs are identifiers and may remain in redacted connection config.
  • A generated URL containing credentials is split. Paperclip stores a safe URL for display/routing and vaults the complete URL. The gateway verifies that the secret URL still matches the public URL before use.
  • Personal credentials live on a user grant and user-scoped secret rows. Organization credentials live on the connection/default organization grant. Choosing "Just me" must never first create or silently fall back to a shared organization credential.
  • Failed setup removes newly created orphan secrets. Removal/revocation clears OAuth state, grant credentials, gateway access, and owned secret material; shared secrets used by another consumer are retained.

Never put raw credentials in any of these places:

  • AppDefinition JSON or the ingestion script
  • connection config or transportConfig
  • application metadata
  • agent, project, or routine plain environment values
  • test fixtures committed to git
  • issue comments, activity details, screenshots, traces, HAR files, or console output
  • PR descriptions or live-proof evidence

Tests must assert absence, not merely avoid printing a secret during the happy path.

Current Default Access Policy

The current product behavior is encoded by recommendedDefaultsForApp in packages/shared/src/app-definitions.ts:

  • Every discovered action is enabled during successful setup.
  • Every active action defaults to Allowed, including write and destructive actions, for every connection method.
  • Permanently blocked provider actions stay disabled.
  • Provider/schema-specific changed-tool quarantine remains a separate catalog concern; do not turn writes Off as a substitute for correct risk classification.

This is an opt-in restriction model. Finishing a connection is still limited to a board user with connection-configuration access, commits the selected action IDs to an auditable profile, and leaves Ask first available for any action. The open default changes the initial policy; it does not create a route around a policy the operator has applied.

If a destructive provider cannot be safe with those defaults, add a narrowly reviewed provider policy with tests. Do not hide a dangerous tool by misclassifying it as read, and do not silently change global defaults in a provider PR.

Golden-Path Agent Tutorial

This is the implementation sequence an autonomous coding agent should follow.

Phase 0: Establish scope and preserve the worktree

  1. Read AGENTS.md, doc/GOAL.md, doc/PRODUCT.md, doc/SPEC-implementation.md, doc/DEVELOPING.md, and doc/DATABASE.md.
  2. Read this runbook, the connections overview, and the security threat model.
  3. Inspect git status --short. Existing changes belong to the user or another task. Do not reset, rewrite, or format unrelated files.
  4. Decide whether the request is research-only, definition-only, or a new runtime capability. A store card is not evidence of runtime support.
  5. Write acceptance criteria before editing. At minimum: actionable store card, setup completes, credentials are vaulted, tools list, one safe read runs, refresh/reconnect works, revoke blocks use, and no secret appears in API or logs.

Phase 1: Research the provider from primary sources

Use current official provider documentation and live protocol metadata. Search results and Vercel captures are leads, not authority.

Record this evidence:

  • Official product/docs URL and date verified.
  • Exact MCP endpoint, including path and trailing-slash behavior.
  • Whether the endpoint uses Streamable HTTP or an older SSE path.
  • Unauthenticated response status and WWW-Authenticate challenge.
  • RFC 9728 protected-resource metadata URL.
  • RFC 8414/OIDC authorization-server metadata URL and exact issuer.
  • Authorization, token, registration, and revocation endpoints when published.
  • PKCE support and token endpoint auth methods.
  • Whether CIMD or DCR is actually advertised.
  • Required scopes. Separate documented minimum scopes from the full discovery list.
  • Access/refresh lifetime, rotation behavior, and terminal refresh errors.
  • Redirect URI constraints: HTTPS, loopback HTTP, exact callback allowlisting, or reviewed-client requirements.
  • Normal prerequisites: account, paid plan, tenant feature flag, administrator consent, preview enrollment, region, project/site identifier.
  • Whether Paperclip itself needs provider approval. Customer-admin approval is self-serve; provider approval of Paperclip is not.
  • Tool/action inventory, provider annotations, known destructive actions, and resource boundaries.
  • Revocation procedure and whether a provider endpoint exists.

Safe research may fetch public metadata, but it must not perform dynamic client registration. Paperclip's catalog preflight is intentionally non-registering:

GET /api/companies/:companyId/tools/apps/:galleryKey/preflight?methodKey=<method-key>

Registration and consent happen only after an explicit Connect action. If the provider requires Paperclip approval or redirect allowlisting that a customer cannot complete, retain the research entry with an unavailable reason and do not expose a connect action.

For researched self-serve MCP providers, update the durable evidence ledger in packages/shared/src/self-serve-mcp-research.json and the dated program plan when eligibility or endpoints change.

Phase 2: Choose the lightest valid product shape

Use this decision tree:

flowchart TD
    A["Provider connection request"] --> B{"Official remote MCP server?"}
    B -->|Yes| C{"Can common auth and fields represent it?"}
    C -->|Yes| D["Curated AppDefinition"]
    C -->|No| E{"One reusable protocol capability?"}
    E -->|Yes| F["Extend common connection runtime plus fixtures"]
    E -->|No| G["Provider wrapper or plugin"]
    B -->|No| H{"Existing generic execution adapter?"}
    H -->|No| I["Do not ship a nonfunctional REST card"]
    H -->|Yes| J["Use the adapter through normal connection governance"]
    G --> K{"Needs UI, tables, workers, migrations, or webhooks?"}
    K -->|Yes| L["Plugin that provisions normal connections"]
    K -->|No| F

Default to an AppDefinition for hosted remote MCP. Use a plugin only for custom product surfaces, tables, workers, migrations, ingestion loops, webhooks, or other real code ownership. A plugin still provisions normal connections and cannot bypass secrets, grants, profiles, policy, gateway, or audit.

Phase 3: Design methods and the setup experience

For every real user choice, create a separate method. Do not create methods for choices Paperclip can infer.

Good separate methods:

  • US OAuth versus EU API-key endpoints.
  • Read versus write capability profiles when the provider publishes distinct servers or scope sets.
  • Standard browser sign-in versus a materially different API-key path.
  • Distinct provider modes such as Postman's minimal, code, and full catalogs.

Avoid separate methods for:

  • DCR versus CIMD. Paperclip chooses automatically.
  • DCR versus a customer-owned OAuth app when both reach the same endpoint. Keep browser sign-in recommended and fold "use your own OAuth app" under Advanced.
  • Optional project filters, read-only switches, or response tuning. These are advanced fields with safe defaults.

The default setup screen should ask only for information required to make the connection work or enforce a real tenant boundary. Follow these rules:

  • Put optional narrowing in fields marked advanced: true.
  • Give hidden fields a defaultValue; never create a hidden required field the server cannot fill.
  • Use setupPrerequisite for steps that must happen before credentials or consent, such as preview enrollment or making a storefront public.
  • Put account/plan/admin limitations in warnings and guidanceMd.
  • Give every method a meaningful label, whenToUse, exact endpoint, risk tier, and official consoleLinks.
  • Use capabilityProfile for user-facing mode names. Do not infer defaults from array order when one mode is the useful write-capable choice.
  • Use grantKinds: ["user"] for providers that only support personal delegated identity.
  • Use requiredResourceFilters as reviewed policy metadata, but remember a label is not enforcement. The provider, gateway, wrapper, or managed header/ query projection must enforce the boundary.

Connector-provided skills and tools

Connectors may contribute bundled skills with optional native tools. Keep provider-specific instructions out of the universal Paperclip skill and provider-specific tools out of the universal runner catalog. Use the trusted connector contribution registry in server/src/services/connector-runtime.ts; AgentMail is the first consumer. This registry describes bundled server implementations, not executable code or skill URLs supplied by a credential or external message.

For each contribution, declare its connector key, bundled skill, namespaced tool definitions, resource-assignment resolver, and execution handler. Use names such as agentmail_send rather than extending core tools with provider-specific branches. Existing MCP connectors continue to use their normal MCP tool catalog; they do not need a duplicate native wrapper just to supply a skill.

Resolve eligibility from current assignments and access. An AgentMail account credential alone does not give an agent email capabilities. An active inbox assigned to that agent does, provided both the inbox connection and saved credential access remain authorized and the experimental chat-connector flag is on. Other connectors must define an equally concrete assignment rule. Keep every lookup company-scoped. Revoked grants, disabled connections, removed assignments, and experimental gates must remove the contribution. Fail closed on lookup errors.

Install skills transparently through the existing runtime skill path. Merge system-managed contributions with the agent's chosen skills for each run, without writing them into its saved skill preferences. Deduplicate multiple resources from the same connector into one skill. Include only authorized resource context, never provider secrets; treat resource values as data. Supply the short skill description for discovery and keep detailed instructions in the skill. The same resolved set must reach local CLI adapters, sandbox adapters, and native runners. Adapters with isolated skill delivery receive the bundle. Adapters that install into shared user directories receive the same assigned skill in the run prompt, including resumed turns, without writing connector files into that directory. Manual skill-sync operations must also exclude automatic connector bundles. The agent Skills page should identify automatic contributions and explain that assignment controls them; they are not independently enabled/disabled there.

Bind tools to the same resolved skill assignment. Native sessions advertise only contributions present in their pinned runtime skill bundle. Include skill content, resource assignments, and tool revisions in session compatibility so a changed assignment cannot reuse stale declarations. Revalidate live assignment, company/task/run authority, and configured action policy on every execution. Removing a tool from discovery alone is not revocation enforcement. Retained provider sessions and previously issued calls must fail after access is revoked.

Avoid shared runtime contamination. Do not install assignment-specific skills into a company-wide or user-wide runtime home. Use immutable skill bundles and scoped runtime directories. Codex CLI connector runs use a separate home per agent and connector-skill revision, seeded from the selected model credential home. Disconnecting returns to a runtime without those skills; another agent must never inherit them. Preserve explicit model identity and normal session recovery.

Required tests cover no assignment, credential access without a resource, authorized assignment, multiple resources with one skill, cross-company access, revocation during a retained run, disabled flags/connections, and reassignment. Verify skill installation and removal in both CLI/sandbox and native execution, including tool discovery, runtime cache changes, and absence of provider secrets. Exercise an actual connector operation through the contributed tool, not just its declaration. Record which runtime paths were tested live versus deterministically.

Connection UX and user journeys

Design the whole journey, from finding the app to doing useful work with an agent. A successful credential exchange is only one step. Describe who the user is, where they start, what they want to accomplish, and where they will see the result. Walk through first use, returning use, and recovery from a failed action. For messaging connections, cover both agent-initiated work and incoming messages that start or continue work.

Separate connecting from assigning an agent a resource. First configure who can use the connection and authenticate with the provider. If the feature also assigns a resource to a specific agent, offer a second wizard from the connection's Permissions view after the connection is saved. Give its entry point a prominent, concrete action name. For example, AgentMail uses “Give an agent an email address,” followed by Agent → Email address → Review. Reuse the saved credential; do not ask for the API key again. Use the existing numbered step pattern, sensible defaults, Back and Cancel, and a clear completion state. Do not add a second wizard when there is no separate assignment to configure.

Let the operator search eligible company agents, including agents not yet on the connection's allowed list. When assigning a resource also grants connection access, make that consequence clear and persist the grant through the existing access machinery. Respect the operator's authority to grant access, and show the selected agent's avatar and name.

Use the minimum text needed to make the next action clear. Prefer familiar controls and precise labels over explanatory paragraphs. Remove repeated headings, redundant access summaries, implementation details, and reassurance that does not help the user decide or act. Keep necessary warnings, meaningful consequences, and actionable errors. Put optional expert settings under a collapsed Advanced disclosure. Link to provider-owned administration, such as AgentMail allowlists, rather than rebuilding it in Paperclip.

Keep ongoing interactions in Paperclip tasks. Connections are where users set up access and configuration; tasks are where they work with agents. Design what happens after setup: how an agent invokes the connection, where incoming work lands, how follow-ups stay associated with that work, and how users see success or recover from failure. Avoid introducing a separate mailbox or provider dashboard as the primary interaction surface.

Use rich cards in the task feed when they make external activity easier to understand. An email card, for example, can show the sender, recipients, body, attachments, and delivery state. Keep external activity distinguishable from internal discussion; a task comment or agent progress update must not imply that an external action occurred. Reuse existing task-feed components and preserve one visible record per external event.

Make interactive Storybooks for setup and actual use. Include the catalog card, access and credential steps, any agent-resource wizard, and the task journeys after setup. Provide a clickable walkthrough plus focused stories for important steps, loading, errors, and recovery. Use realistic fixtures and clearly label simulated actions. Reuse production components as implementation lands, and replace obsolete stories so the examples describe the current experience. Storybooks support design review and deterministic interaction tests; they do not replace a real-provider browser test.

Phase 4: Add official branding before exposing the app

Every store-visible provider needs an official local mark. A letter tile is only a runtime image-failure fallback.

  1. Find the provider's official brand kit, product site, or official repository.
  2. Prefer an official SVG. Use a high-resolution transparent PNG only when no official SVG is available.
  3. Do not use Google's favicon proxy, scrape a random icon site, or generate an imitation.
  4. Sanitize SVGs. Reject scripts, foreignObject, event-handler attributes, external executable content, or unsafe references.
  5. Save assets under ui/public/brands/apps/. Add a -dark variant only when the normal mark loses contrast in dark mode.
  6. Add the provider to ui/public/brands/apps/manifest.json with slug, local asset, optional dark asset, official source URL, exact upstream asset URL, asset type, visibility, and dark-variant requirement.
  7. Let the ingestion script derive branding.logoUrl and darkLogoUrl from the provenance manifest.

The manifest test decodes PNG headers, requires at least 128 by 128 pixels, sanity-checks SVG markup, verifies files exist, and requires store-visible definitions and visible provenance entries to match exactly.

Phase 5: Author the definition at the durable source

The checked-in provider JSON files are generated. Do not edit one and stop.

  1. Add or update the provider in scripts/ingest-app-definitions.mjs.
  2. Update packages/shared/src/self-serve-mcp-research.json when it belongs to that program.
  3. Add branding provenance and assets first; generation fails closed when branding is missing.
  4. Regenerate definitions:
pnpm connections:ingest-app-definitions

The default ingestion corpus is the Vercel research checkout at ../../paperclip-content/research/connections/vercel/templates. Override it when necessary:

PAPERCLIP_CONTENT_TEMPLATES=/absolute/path/to/templates \
  pnpm connections:ingest-app-definitions

Generation currently validates the 99-capture corpus and rewrites provider JSON, the generated TypeScript registry, and the ingestion report. A PR must contain the human-authored source and generated output. Inspect the diff after generation; do not accept unrelated provider churn.

Minimal automatic OAuth example:

{
  "key": "mcp-oauth",
  "label": "Sign in with Example",
  "transport": "mcp_remote",
  "auth": "oauth",
  "ownershipModes": ["dcr"],
  "whenToUse": "Use browser sign-in for the hosted MCP server.",
  "defaults": {
    "serverUrl": "https://mcp.example.com/mcp",
    "scopesHint": ["example.read", "example.write"]
  },
  "guidanceMd": "Connect the workspace agents should use.",
  "consoleLinks": {
    "docs": "https://docs.example.com/mcp"
  },
  "riskTier": "S3"
}

Minimal customer OAuth example:

{
  "key": "mcp-own-oauth",
  "label": "Use your own OAuth app",
  "transport": "mcp_remote",
  "auth": "oauth",
  "ownershipModes": ["customer"],
  "whenToUse": "Register an OAuth app, then enter its client ID and secret.",
  "defaults": {
    "serverUrl": "https://mcp.example.com/mcp"
  },
  "guidanceMd": "Register Paperclip's callback URI in the provider console.",
  "consoleLinks": {
    "register": "https://example.com/developers/apps",
    "docs": "https://docs.example.com/mcp/oauth"
  },
  "riskTier": "S3"
}

Minimal API-key example:

{
  "key": "mcp-api-key",
  "label": "Use an API key",
  "transport": "mcp_remote",
  "auth": "api_key",
  "ownershipModes": ["customer"],
  "whenToUse": "Use a restricted key from the provider console.",
  "defaults": {
    "serverUrl": "https://mcp.example.com/mcp"
  },
  "credentialFields": [
    {
      "key": "authorization",
      "label": "Example API key",
      "type": "password",
      "required": true,
      "placeholder": "ex_...",
      "secret": true
    }
  ],
  "keyPlacement": {
    "location": "header",
    "name": "Authorization",
    "prefix": "Bearer "
  },
  "guidanceMd": "Create a key limited to the resources agents need.",
  "riskTier": "S3"
}

Generated-URL example:

{
  "key": "generated-url",
  "label": "Paste generated MCP URL",
  "transport": "mcp_remote",
  "auth": "none",
  "ownershipModes": ["customer"],
  "whenToUse": "Paste the complete server URL generated by the provider.",
  "defaults": {},
  "guidanceMd": "Create a server in the provider, then paste its URL.",
  "riskTier": "S3"
}

No-auth tenant-template example:

{
  "key": "public-mcp",
  "label": "Public storefront",
  "transport": "mcp_remote",
  "auth": "none",
  "ownershipModes": ["customer"],
  "whenToUse": "Connect a public tenant endpoint.",
  "defaults": {
    "serverUrlTemplate": "https://{tenantDomain}/api/mcp"
  },
  "tenantFields": [
    {
      "key": "tenantDomain",
      "label": "Tenant domain",
      "type": "text",
      "required": true,
      "placeholder": "store.example.com",
      "validation": {
        "pattern": "^[A-Za-z0-9.-]+$",
        "maxLength": 255
      }
    }
  ],
  "guidanceMd": "Enter the permanent public tenant domain.",
  "riskTier": "S2"
}

Use defaults.toolArgumentDefaults only for required, provider-documented protocol metadata that Paperclip owns, not to force a user's business input. Managed arguments are deep-merged after caller input and win on collisions; the same fields are removed from the agent-visible and Test-tab input schema.

Phase 6: Add generic runtime support only when needed

Before adding code, prove the manifest cannot express the provider.

Common extension points:

  • packages/shared/src/types/app-definition.ts and packages/shared/src/validators/app-definition.ts for a reusable manifest capability.
  • normalizeConnectionMethodConfig for validated tenant/extension fields and header/query projection.
  • projectedConnectionHeaders, projectedConnectionToolArguments, and projectedConnectionToolInputSchema for server-managed request material.
  • OAuth discovery/client/token/refresh functions in server/src/services/tool-access.ts.
  • MCP invocation in server/src/services/tool-gateway.ts.
  • classifyRisk only for reviewed provider exceptions that generic annotations and name classification cannot represent safely.
  • ConnectionSetupFlow only when a schema-driven setup capability genuinely cannot render the flow.

When adding a reusable field:

  1. Update shared TypeScript types.
  2. Update the Zod validator with cross-field invariants.
  3. Update server normalization and invocation projection.
  4. Update UI rendering and request types.
  5. Add fixture tests for valid, invalid, redacted, reconnect, and invocation behavior.
  6. Document the new field here.

Keep ConnectToolApp synchronized across UI and server. Never create a UI-only request shape that drops connectionMethodKey, oauthClient, grantKind, or credential source data.

Phase 7: Write deterministic tests before using a real account

At minimum, add or update tests in these layers:

Manifest

  • Schema validates.
  • Slug and method keys are unique and stable.
  • Endpoint, scopes, ownership modes, risk tier, prerequisite, and field placement equal the reviewed values.
  • Store visibility matches the intended rollout state.
  • Official local branding and provenance exist.
  • Hidden fields have defaults; required fields have placeholders.
  • The default setup path asks only for truly required configuration.

Server/service

  • Method selection is required when multiple methods are real choices.
  • Tenant fields normalize, validate, and reach the exact URL/header/query.
  • Unknown config fields are rejected.
  • API keys/client secrets/tokens become encrypted refs and never appear in the response.
  • Failed setup removes newly created secrets and draft rows when appropriate.
  • DCR, CIMD, and manual-client paths use fixture metadata and never require a real provider.
  • Scope widening beyond scopesHint is rejected.
  • OAuth state, actor/session, issuer, redirect, resource, and company bindings fail closed.
  • Refresh rotates safely; terminal invalid_grant requires reauthorization.
  • Health, catalog refresh, reconnect, removal, and secret cleanup work.
  • SSRF, private/link-local address, redirect, header-name, and header-value protections remain intact.
  • Company A cannot observe or invoke company B's connection.
  • Managed arguments/headers cannot be spoofed by caller input.
  • Risk exceptions classify every reviewed tool correctly.

UI

  • Browse has an actionable route for an available capability-backed definition.
  • Hidden/unavailable providers do not show a dead Connect button.
  • Direct /apps/connect?source=<slug> opens the selected provider, not the generic gallery.
  • Automatic OAuth, customer OAuth, API key, generated URL, no-auth, required tenant field, prerequisite, warning, and Advanced disclosures render as declared.
  • Finish setup resumes the exact draft using resumeConnectionId.
  • Optional customer OAuth details stay folded when automatic OAuth exists.
  • Setup success leads to the connection's Test page, or to Permissions when a separate agent-resource assignment is the next step. Follow the connection UX guidance.
  • Interactive Storybooks cover setup and ongoing task interactions, including relevant failure states; the walkthrough matches the implemented journey.
  • Missing images fall back at runtime, while manifest acceptance still fails missing branding.

Useful focused command:

pnpm exec vitest run \
  packages/shared/src/app-definitions.test.ts \
  server/src/__tests__/tool-access-service.test.ts \
  server/src/__tests__/generic-mcp-connection.test.ts \
  server/src/__tests__/tool-connection-removal.test.ts \
  ui/src/pages/apps/AppsConnect.test.tsx \
  ui/src/pages/apps/Browse.test.tsx

Use -t '<provider or behavior>' while iterating, then run each affected file without a test-name filter before handoff.

Targeted type checks:

pnpm --filter @paperclipai/shared typecheck
pnpm --filter @paperclipai/server typecheck
pnpm --filter @paperclipai/ui typecheck

If UI code changed, also run:

pnpm check:token-gates

Phase 8: Start an isolated instance and verify the setup UI

Use a worktree-local instance; never point two worktrees at the same embedded database.

paperclipai worktree init
pnpm dev

Confirm the actual port with pnpm dev:list and verify health:

curl -fsS http://localhost:<port>/api/health
curl -fsS http://localhost:<port>/api/companies

Walk the user path:

  1. Open /<company-prefix>/apps.
  2. Confirm branding, copy, visibility, ordering, and Connect state.
  3. Open /<company-prefix>/apps/connect?source=<slug> directly.
  4. Verify prerequisites and warnings appear before credentials/consent.
  5. Exercise every method. Do not test only the default method.
  6. Cancel or interrupt once. Confirm the store shows Finish setup and that it returns to ?source=<slug>&resume=<connection-id> without creating another draft.
  7. Complete setup. Confirm the connection is active/healthy and opens /<company-prefix>/apps/<connection-id>/permissions, then use the action's Test button.

For OAuth, the instance callback must be browser-reachable and must match the provider registration. Loopback HTTP is acceptable only when provider and Paperclip redirect policies permit it. Browser-started setup on an authenticated private instance automatically uses the same-origin HTTPS address that served the setup page, including a Tailscale Serve address; the request must pass the hostname and board-mutation guards. An explicit PAPERCLIP_PUBLIC_URL remains available for non-browser starts and unusual proxy topologies. Internal service hostnames are not valid browser callback origins.

Use the browser signed-in session only for an explicitly authorized live proof. Do not inspect cookies, storage, saved passwords, or unrelated account data.

Phase 9: Perform the real-provider proof

Deterministic fixtures prove Paperclip logic. A store-ready provider also needs one account-bound proof for every method being exposed.

Run this exact lifecycle:

  1. Preflight — public metadata only; no registration or credentials.
  2. Connect — finish provider consent or enter the credential.
  3. Catalog — list tools and compare them with reviewed expectations.
  4. Safe read — execute one narrow, non-mutating action in the Test page.
  5. Write classification — confirm known writes/destructive actions appear in the correct risk group and default policy.
  6. Agent path — run one action through an actual agent/run-scoped gateway, not only the board Test helper, when the connection changes gateway logic.
  7. Refresh/reconnect — refresh the catalog, reconnect or force a safe token refresh, and repeat the safe read.
  8. Revoke/remove — revoke at the provider or remove in Paperclip. Confirm tools disappear or calls fail closed immediately.
  9. Reconnect after removal — when supported, confirm the retained identity and history are reused rather than duplicated.
  10. Secret inspection — inspect API responses, application logs, activity, audit, screenshots, and evidence artifacts for the exact canary credential. It must be absent.

Evidence should record only:

  • provider and method key
  • date, environment, endpoint origin/path, and connection ID when non-sensitive
  • catalog tool names/counts and schema hashes
  • policy/risk result
  • redacted success/failure codes
  • revoke/reconnect outcome

Do not record token values, secret-bearing URLs, authorization codes, provider session details, personal email, tenant content, HAR files, or pre-callback screenshots containing provider data.

When consent UI has an intentional delay or requires a real pointer/keyboard interaction, honor that behavior before classifying it as failure. Keep provider consent problems separate from callback, token exchange, catalog, and gateway failures.

Phase 10: Verify APIs and audit without exposing credentials

Useful read-only endpoints after setup:

GET /api/companies/:companyId/tools/connections
GET /api/tool-connections/:connectionId
GET /api/tool-connections/:connectionId/catalog
GET /api/tool-connections/:connectionId/activity?limit=50
GET /api/tool-connections/:connectionId/grants

Mutating verification endpoints:

POST /api/tool-connections/:connectionId/health-check
POST /api/tool-connections/:connectionId/catalog/refresh
POST /api/tool-connections/:connectionId/test-calls
DELETE /api/tool-connections/:connectionId

Responses may contain secret IDs, version selectors, header names, prefixes, scope names, expiry timestamps, and redacted provider metadata. They must not contain secret values. Logs and activity should identify the operation and outcome without echoing provider-authored credential-bearing errors.

Phase 11: Run the PR-ready verification ladder

Run the smallest relevant suite first, then the full ladder when the change is ready for review:

pnpm check:token-gates
pnpm -r typecheck
pnpm test:run
pnpm build

Also run an affected Apps browser suite when routing, setup, OAuth popup, finish-setup, Test page, or branding behavior changed:

pnpm test:e2e

Do not hide failures. Classify each as introduced, pre-existing, environmental, or live-provider-only, and include the exact command and result in the PR.

Before staging:

git diff --check
git status --short
git diff --stat
git diff -- \
  scripts/ingest-app-definitions.mjs \
  packages/shared/src/app-definitions \
  packages/shared/src/app-definitions.generated.ts \
  ui/public/brands/apps \
  server/src/services/tool-access.ts \
  server/src/services/tool-gateway.ts

Review every generated change. Verify no credential, provider account data, unrelated worktree change, or temporary evidence file is staged.

Phase 12: Prepare and submit the pull request

  1. Keep commits scoped. A typical split is definition/branding, generic runtime capability, and tests/docs. Do not split generated output from its source.
  2. Do not commit pnpm-lock.yaml; GitHub Actions owns it in this repository.
  3. Read .github/PULL_REQUEST_TEMPLATE.md immediately before writing the PR body.
  4. Fill every required section:
    • Thinking Path — why this is a catalog entry, chosen transport/auth, research evidence, and why no lighter path works.
    • What Changed — manifest, branding, runtime, UI, tests, and docs.
    • Verification — deterministic commands plus sanitized live proof.
    • Risks — scopes, provider preview/admin gates, token behavior, catalog drift, destructive actions, and rollback/de-list plan.
    • Model Used — provider, exact model ID, context window, and relevant capabilities, or the template's human-authored value.
    • Checklist — every item checked truthfully.
  5. In the PR, link official provider docs and exact metadata endpoints. Do not link only to search results or third-party tutorials.
  6. Mark live proof that was not run as outstanding; never equate a mocked OAuth test with provider validation.
  7. Do not merge as part of connection authoring unless the task explicitly authorizes merging.

Suggested PR verification block:

## Verification

- `pnpm exec vitest run <focused files>` — passed
- `pnpm check:token-gates` — passed
- `pnpm -r typecheck` — passed
- `pnpm test:run` — passed
- `pnpm build` — passed
- Live `<provider>/<method>` proof on `<date>`:
  connect ✓, list tools ✓, safe read ✓, refresh/reconnect ✓, revoke ✓,
  secret scan ✓

AppDefinition Field Reference

App-level fields

Field Meaning and rule
schemaVersion Must be 1. Change only with a versioned migration plan.
slug Stable lowercase kebab-case identity. Never rename after connections exist without a migration.
name Provider/product name shown to users.
description Plain-language outcome, not protocol marketing.
categories One or more supported catalog categories.
featured Optional merchandising signal, not availability.
branding Local official logoUrl, optional darkLogoUrl; ingestion derives this from provenance.
urlPatterns HTTPS patterns used to recognize pasted/generated provider URLs. Keep narrow enough to reject lookalikes.
docsUrl Current official setup/protocol docs.
setupPrerequisite A prerequisite users must understand or complete before credentials/consent. Includes CTA and optional ordered steps.
redirectConstraints Currently https-or-loopback-http; fail before provider navigation when violated.
methods Every genuinely supported connection method. At least one.
availability Instance/provider availability and user-facing reason. Unavailable entries must not expose a dead action.
ownershipAvailability Deployment override for ownership modes. Defaults currently enable customer and dcr, disable platform modes.

Method fields

Field Meaning and rule
key Stable method key stored on the connection as connectionMethodKey.
label User-facing method label. Required in practice when multiple methods exist.
transport mcp_remote, local_stdio, or specialized rest_api.
auth oauth, api_key, or none.
ownershipModes Allowed OAuth client ownership modes; also present for non-OAuth customer configuration.
oauthStrategy Managed broker strategy. New definitions use paperclip_cloud_connector; paperclip_id_connector is recognized only to require migration when an old grant expires. The protocols and provider clients are not interchangeable. Only valid for OAuth.
connectorProfile Managed connector capability/scope profile, required with oauthStrategy.
capabilityProfile User-facing read/write/mode grouping used for method selection.
grantKinds Restricts identity to organization and/or user; omit for flexible methods.
whenToUse One sentence distinguishing this method from alternatives.
defaults.serverUrl Exact fixed endpoint. For discovery-capable OAuth, omit fixed auth endpoints.
defaults.serverUrlTemplate HTTPS endpoint with placeholders supplied by declared tenant/extension fields. Mutually exclusive with serverUrl.
defaults.discoveryUrl Provider-specific discovery override only when reviewed metadata requires it.
defaults.authorizationEndpoint / tokenEndpoint Authoritative fixed endpoints. A complete pair bypasses discovery, so ship them only when the provider lacks trustworthy discovery.
defaults.metadataUrl Authorization metadata hint.
defaults.scopesHint Explicit reviewed allowlist. Omit scope when docs do not require one; never copy every discovered scope.
defaults.oauthAuthorizationParams Reviewed access_type=offline and/or prompt=consent behavior.
defaults.toolArgumentDefaults Server-owned provider protocol arguments, hidden from caller schemas and authoritative on collision.
tenantFields Account/project/region/resource fields. Keep only required boundaries visible by default.
extensionFields Additional method-specific configuration rendered by the same common form.
configRequirements.atLeastOneOf Requires one of named tenant/extension fields. Every key must exist.
credentialFields Write-only secret/non-secret credential inputs. API-key methods require them in practice.
keyPlacement Provider request placement. Remote MCP should use the proven header path unless a new projection is implemented and tested.
guidanceMd Setup and scoping guidance. No secrets or internal environment-variable names.
consoleLinks Official registration, key, settings, and docs destinations.
warnings Plan, preview, admin, financial, production-data, or destructive-action caveats.
variants Legacy/simple variant metadata. Prefer explicit methods plus capabilityProfile for materially different endpoints/auth.
riskTier S1-S4 provider/method sensitivity used for review and validation.
requiredResourceFilters Reviewed resource boundaries. Must be backed by enforcement, not only copy.
credentialSources.vercelConnect Reviewed services, principal modes, scopes, and header projection for the Vercel exception.

Field definition rules

FieldDef supports text, password, textarea, datetime, select, and checkbox.

  • A required non-checkbox field needs a placeholder.
  • A select needs at least one option.
  • A hidden field needs a default value.
  • Use secret: true only for write-only credential input; tenant/extension config must never smuggle secrets into connection config.
  • advanced: true keeps optional expert configuration folded.
  • validation.pattern is a JavaScript regular expression string; also set a practical maxLength.
  • transport.location: "query" modifies the normalized server URL.
  • transport.location: "header" creates a managed, validated non-secret configuration header. It is separate from keyPlacement, which projects a vaulted credential.
  • CSV fields deduplicate comma/newline-separated values.
  • omitFalse prevents a false checkbox from adding a query/header value.

Troubleshooting And Failure Classification

Symptom Likely layer What to inspect
Store says Coming soon or Connect route is dead Definition/availability/routing CONNECTABLE_APP_SLUGS, store hidden set, method capability checks, Browse tests.
Direct source link shows generic connection chooser UI route state AppsConnect, ConnectionSetupFlow, source slug lookup, availability.
Finish setup opens Edit config and cannot continue Draft identity/resume resumeConnectionId, stored sourceTemplateKey, connectionMethodKey, exact draft status.
OAuth never redirects Method capability/client resolution ownership modes, metadata discovery, callback origin, manual-client requirement.
Provider rejects redirect URI Deployment/provider rule actual browser origin, PAPERCLIP_PUBLIC_URL, redirectConstraints, provider app registration.
OAuth succeeds then connection needs reconnect Grant/secret sync or refresh organization versus user grant, token refs, default grant sync, expiry/refresh lease, invalid_grant.
Tools list but calls return 401 Token audience/scope/placement RFC 8707 resource, scopesHint, header prefix/name, provider endpoint path.
Health works but Test call fails Gateway projection/policy selected grant, managed headers/arguments, effective profile/policy, catalog entry risk/status.
Required provider boilerplate appears in Test Managed schema projection toolArgumentDefaults and projectedConnectionToolInputSchema.
API key saves but is not sent Unsupported placement or missing ref credentialFieldsFor, keyPlacement, connection/grant refs, gateway header resolution.
Config asks for project ID the provider does not require Manifest UX make it optional/advanced, add default, or remove it; test the zero-config path.
Connected card still says Connect Identity matching application/source slug, retained app status, connection-to-definition association.
New tool is classified read Risk inference annotations, namespaced/camelCase verb normalization, provider exception set, fixture.
Connection from another company is visible Authorization bug stop; add company-scope negative tests before any further live testing.
Raw credential appears anywhere Security incident stop, revoke/rotate it, remove evidence, trace every response/log/audit path, add a canary regression test.

When a bug appears on one provider, first reproduce it with a fixture or a second provider of the same auth/transport type. Fix the shared path when the failure is generic. Keep provider workarounds narrow and documented.

Definition Of Done

A catalog connection is ready only when every applicable item is true:

  • Official docs and live metadata agree on the endpoint and auth flow.
  • Self-serve/provider-approval classification is documented.
  • Every exposed method has completed the full live lifecycle.
  • The default path asks only for required configuration.
  • Optional own-OAuth and expert fields are folded under Advanced.
  • Scopes are explicit and contained; caller widening is rejected.
  • Official local branding and provenance pass validation in light and dark themes.
  • Ingestion source and generated definitions are synchronized.
  • Credentials and tokens are stored only as encrypted/external refs.
  • Shared and personal grant semantics are correct.
  • Health and catalog discovery succeed.
  • One safe read succeeds in the Test page and, where relevant, an agent run.
  • Writes/destructive tools are correctly classified and use current tier defaults.
  • Refresh, reconnect, revoke/remove, and post-removal reconnect are verified.
  • Company isolation, SSRF, OAuth binding, redaction, and cleanup tests pass.
  • API responses, logs, activity, and evidence contain no credential values.
  • Focused tests, token gates, typecheck, test suite, and build pass or exact blockers are reported.
  • PR uses every section of the repository template and links official evidence.

Detailed Design Checklist

The golden-path tutorial above is the operational sequence. The following steps are the design-review checklist: use them when the connection proposal needs a more formal transport, credential, action, and governance analysis.

Step 1: Confirm It Is A Catalog Entry

Default to a catalog entry when the vendor can be represented as metadata plus a transport:

  • The connection points at a remote MCP endpoint, an approved local stdio template, or a generated shim over a documented API.
  • The setup flow only needs normal fields, OAuth redirect handling, resource filters, policy defaults, and health/catalog checks.
  • The vendor does not need its own database tables, background workers, custom issue-thread interactions, or dedicated UI pages.

Use a plugin only when the integration needs code that cannot fit inside the common connection model:

  • Product surface: custom pages, dashboards, panes, or rich configuration UI beyond schema-driven forms.
  • Data model: plugin-owned tables, migrations, or long-lived local state.
  • Execution: workers, schedulers, webhooks, sync loops, file processors, or vendor-specific runtimes that are not a simple transport shim.
  • Packaging: a third party wants to ship the integration as an extension package.

A plugin may bundle one or more catalog entries, but it must still create normal applications, connections, credential refs, catalog entries, profiles, policies, and audit events. Plugin code must not bypass the gateway, policy engine, company_secrets, changed-action quarantine, or call-event audit log.

Step 2: Classify The Reuse Path

Classify the vendor before writing metadata. Use the PAP-2432 matrix terms so rollout planning, security review, and QA can compare providers consistently.

Reuse path Use when Typical transport Examples from the matrix
MCP-direct The vendor exposes an official or stable MCP server whose tools map cleanly to Paperclip grants. mcp_remote; local_stdio only for approved trusted templates. Linear, Notion, Sentry, Vercel, Exa, Apify, Context7.
OpenAPI-shim The vendor has a documented REST/OpenAPI surface but no stable MCP server, and a generated/thin shim can expose safe actions. Shim service or approved template that presents an MCP-compatible catalog to Paperclip. Datadog, Apollo, QuickBooks, Ramp/Brex, Zendesk.
Vendor-deep-wrapper The vendor boundary depends on app-installation tokens, event validation, rich domain semantics, resource grants, or high-risk writes. Vendor-specific wrapper behind the same connection model. GitHub, Slack, Google Workspace writes, Atlassian, Microsoft 365, Cloudflare, Figma, Stripe, Salesforce, HubSpot, Intercom, PagerDuty.

Record the classification in the proposal along with the transport and the reason a lighter path is or is not enough.

Step 3: Pick Auth And Credential Ownership

Choose one method auth mode:

  • OAuth: delegated user or workspace authorization. The OAuth client may come from DCR/CIMD, a customer-created client, a deployment-preconfigured client, or a reviewed Paperclip Cloud connector profile. Do not assume Paperclip owns a shared client registration.
  • API key: operator-supplied token or key. Use only when the provider supports a suitably restricted key and the value is stored as a company_secrets ref.
  • None: public/read-only systems and provider-generated URLs. A generated URL may still contain a secret and must be split and vaulted.

Installation credentials such as bot tokens and GitHub App installation tokens are provider-specific credential shapes. Model them through the generic secret and grant architecture; do not invent an AppDefinition.auth value named app-installation because the current schema accepts only oauth, api_key, and none.

Credentials normally live in company_secrets with redacted metadata and versioned material. The catalog entry records the secret binding shape, not the secret value:

{
  "credentialSecretRefs": [
    {
      "configPath": "credentials.authorization",
      "label": "Linear OAuth access token",
      "required": true
    }
  ],
  "credentialRefs": [
    {
      "name": "Authorization",
      "placement": "header",
      "key": "Authorization",
      "prefix": "Bearer ",
      "secretId": "<resolved at connect time>"
    }
  ]
}

Do not add durable vendor credentials to agent env, project env, runtime env, adapter config, issue comments, screenshots, logs, fixture JSON, or plugin config. Agents receive a run-scoped gateway token; Paperclip resolves the vendor credential server-side and audits the call.

For a Vercel-eligible method, add reviewed credentialSources.vercelConnect metadata: allowed Vercel service identifiers, the app or user principal mode, exact token scopes, and the header placement. This is an allowlist, not a copy of Vercel's connector form. Only authenticated mcp_remote methods qualify. Do not infer Paperclip ownership from Vercel's “Managed” label. A Vercel-backed connection remains customer/dcr according to the existing Paperclip model. The resulting connection has an external connector ref and zero Paperclip credential secret refs; its grants likewise use external metadata or secret refs, never both.

Step 4: Author The AppDefinition

Author an AppDefinition as the canonical data record for the app and every supported connection method. It must explain what the operator gets without exposing protocol details in prosumer surfaces. Developer docs can mention transport, MCP, shim, and gateway terms; the Apps gallery copy should use plain app/action language.

Capture:

  • schemaVersion: 1 and a stable lowercase hyphenated slug, such as linear.
  • name, description, categories, branding, urlPatterns, and docsUrl: user-facing and import metadata. Branding points to vetted local assets under /brands/apps/.
  • methods: explicit combinations of transport (mcp_remote, rest_api, local_stdio), auth (oauth, api_key, none), and ownershipModes (platform_shared, platform_provisioned, customer, dcr).
  • credentialFields plus keyPlacement: labels, write-only value fields, vendor-call placement, header/key name, and prefix. The saved value becomes a company_secrets ref, never a plain env/config value.
  • tenantFields and extensionFields: keep unavoidable identity and resource boundaries in the default flow. Mark optional scope reduction, feature/tool filters, response modes, and transport tuning with advanced: true, and give advanced fields working defaults that do not require the operator to expand the disclosure.
  • defaults: exact server/template URL, optional discovery or OAuth endpoint hints, a contained scopesHint, and only reviewed managed tool arguments.
  • grantKinds, oauthStrategy, connectorProfile, credentialSources, capabilityProfile, variants, configRequirements, and requiredResourceFilters only when their documented semantics apply.
  • setupPrerequisite, warnings, guidanceMd, and consoleLinks: everything the operator must know before credentials or consent.
  • riskTier: the method-level S1-S4 tier used for review and validation.
  • availability: whether the connection is usable on this instance and the precise reason when it is not.

Keep AppDefinition metadata deterministic and company-scoped at install time. Global catalog data names capabilities; company connection and grant rows hold the configured instance, subject/provider tenant, secret refs, resource filters, status, health, and audit history.

Step 5: Model Resource Filters

Every connector proposal needs resource filters before write actions are enabled. Filters are part of the connection configuration and must be enforced by the gateway or wrapper, not only by UI affordances.

Common filter dimensions:

  • Account boundary: workspace, org, team, tenant, site, portal, realm, account.
  • Resource boundary: repo, channel, page, database, project, zone, file, folder, dashboard, issue queue.
  • Object boundary: issue status, labels, branch, environment, object type, record type, field list, attendee domain.
  • Egress boundary: domain allow/deny list, result limits, content category, attachment/file-type limits.
  • Mutation boundary: create-only, draft-only, comment-only, no delete, no external send, dry-run required.

The connection health and catalog discovery steps should fail or warn when required filters are absent for S3/S4 providers.

Step 6: Define The Action Catalog

List each initial action before implementation. Do not rely on vendor tool names alone; Paperclip needs normalized metadata for review, policy, and audit.

For each action, capture:

  • Stable tool name and user-facing title.
  • Description in operator language.
  • Input and output schema.
  • Read/write/destructive flags and risk level.
  • Resource filter fields used by the action.
  • Redaction plan for arguments and results.
  • Expected audit fields.
  • Whether the action is enabled, disabled, or quarantined by default.
  • Negative access case: ungranted actor, disallowed resource, revoked connection, or cross-company attempt.

Risk classes:

Risk Examples Default
read Search, list, fetch metadata/content inside allowed resources. Active when profile includes the app or read risk level.
write Create issue, add comment, update status, append block, trigger redeploy. Allowed under the current new-connection default. Operators may narrow individual actions.
destructive Delete, refund, cancel production deployment, send external message, broad tenant mutation. Allowed under the current new-connection default. A provider with meaningful destructive capability should receive an explicit security review and may receive a narrower provider policy.

Changed-action quarantine is available when a connection sets quarantineNewEntries: true. Use it for providers whose catalog can change without a Paperclip release. This is runtime setup behavior, not currently an AppDefinition field, so adding it to a new curated class requires a shared implementation and tests. Do not claim quarantine in provider copy unless the connection actually enables it.

Step 7: Select The Wizard Path

The wizard path comes from auth mode and transport:

These paths describe authentication and provisioning. Apply the connection UX guidance to the user-facing sequence: choose access before authentication, then configure any per-agent resource through a separate wizard on the saved connection.

Auth mode Operator path Stored result
OAuth Gallery card -> Connect -> vendor consent -> callback -> configure filters -> health/catalog -> access defaults. OAuth token material in company_secrets; connection metadata redacted.
API key Gallery card -> paste key -> configure filters -> health/catalog -> access defaults. Key material in company_secrets; no raw key returned after save.
None Gallery card -> configure allowed resources -> health/catalog -> access defaults. No vendor secret; connection row still carries config and audit scope.

Provider-generated URLs also use auth: "none", but the complete URL is vaulted when it contains credential material. Installation-style providers use the nearest supported auth path plus provider-specific setup guidance; they do not add a fourth manifest auth mode.

The operator should see Apps, Connections, and Review language. Keep protocol language behind Developer/Advanced copy.

Request only the documented scope set the reviewed connection needs; never adopt every scope returned by discovery. Operators should not have to predict every future tool during setup. Keep the default view to the minimum inputs needed for a working connection, fold optional expert controls under one collapsed Advanced disclosure, and enforce execution afterward through Paperclip's resource boundaries, risk classification, tier defaults, optional quarantine, and audit.

Step 8: Apply Governance Defaults

Governance is automatic because every catalog entry becomes a normal tool-access object:

  1. Catalog status gates first: disabled and quarantined deny immediately.
  2. Profiles decide which actors can see catalog entries. Bindings can target company, project, agent, routine, or issue scopes.
  3. Policies decide whether a visible action is allowed, blocked, rate-limited, or requires approval.
  4. Ask-first calls create action requests with signed arguments. Approval applies only to the reviewed argument shape and unchanged schema hashes.
  5. Every decision and call writes audit with actor, run, issue, connection, catalog entry, decision, reason code, redaction summary, outcome, and latency.

Recommended defaults for a new catalog entry:

  • Use the central recommendedDefaultsForApp policy. Do not invent a provider default in UI code.
  • All active actions default Allowed for every method tier. Operators can move individual actions to Ask first or Off after setup.
  • Classify a method S4 when its normal catalog includes payments, external sends, refunds, production deployment, deletion, tenant-wide administration, or comparable high-impact mutations.
  • Add an explicit block only for a tool Paperclip must never expose, and prove it with a provider-specific negative test.
  • Enable changed-tool quarantine for catalogs that can drift independently, and add a rate limit for quota-sensitive or paid APIs.

Step 9: Align With Production Validation

PAP-12373 owns real-vendor gallery smoke evidence and connector validation. Do not duplicate that issue's screenshot/evidence matrix in this playbook. A connector proposal should instead state exactly how it will be validated there:

  • Connect succeeds against the real vendor using production-like OAuth/app/key setup.
  • Catalog discovery produces the expected actions and the declared changed-tool behavior.
  • An allowed read call succeeds through the gateway.
  • A write call is Allowed by the new-connection default unless an explicit provider or operator policy narrows it.
  • A blocked/quarantined action, when declared, cannot be listed or invoked by an agent.
  • Revocation removes tools and blocks execution immediately.
  • Activity/audit rows prove actor, run/issue context, resource id, decision, reason code, and outcome.

If a gallery card cannot pass this path against a real vendor, de-list it or mark it unavailable until the missing auth, transport, or governance dependency is fixed.

MCP-Direct Connections (Hosted MCP + OAuth)

Many vendors now expose an official hosted MCP server whose authorization server is discovered from the MCP endpoint itself, instead of documenting fixed OAuth URLs. For these connectors the manifest's oauth block is a hint at most; the broker resolves endpoints at connect time:

  1. GET <serverUrl> unauthenticated returns 401 with a WWW-Authenticate header naming the protected-resource metadata URL (RFC 9728).
  2. GET /.well-known/oauth-protected-resource[/<path>] names the authorization server(s).
  3. GET /.well-known/oauth-authorization-server (RFC 8414) yields authorization_endpoint, token_endpoint, and — when the vendor supports dynamic registration — registration_endpoint.

For an issuer that has a path, step 3 is tried in both the RFC 8414 insertion form (/.well-known/oauth-authorization-server<path>) and the widely deployed OIDC suffix form (<path>/.well-known/oauth-authorization-server), and a document whose issuer disagrees with the issuer used to build the URL is discarded. Authorization, token, and refresh requests all carry the RFC 8707 resource indicator naming the canonical MCP endpoint, and RFC 9207 iss is validated against the persisted expected issuer when the authorization server returns it.

The broker implements this in discoverOAuthEndpoints (server/src/services/tool-access.ts), but discovery is not unconditional. oauthEndpointsForConnection resolves endpoints in this order:

  1. If the manifest's method defaults ship a complete pair (authorizationEndpoint and tokenEndpoint), those are used unconditionally. discoverOAuthEndpoints never runs in this case, so endpoints stored on the connection's own OAuth config and 401 challenge hints are not consulted at all.
  2. Otherwise, for mcp_remote connections, the broker calls discoverOAuthEndpoints, which first checks endpoints already stored on the connection's own OAuth config (falling back field-by-field to the 401 challenge hints); a complete stored/hinted pair is used as-is — no .well-known fetch.
  3. Only when neither of the above yields a complete pair does the broker run the RFC 9728 → RFC 8414 discovery chain above.

Consequence: complete manifest endpoint hints are authoritative, not hints — they override even endpoints that an earlier discovery persisted on the connection, and if they go stale the broker keeps using them. For discovery-capable vendors, ship only serverUrl in defaults (as notion.json does) so the broker discovers fresh endpoints at connect time; add explicit authorizationEndpoint/tokenEndpoint only for vendors that do not publish RFC 9728/8414 metadata, and then own keeping them current.

Dynamic client registration (RFC 7591)

Vendors whose authorization server advertises a registration_endpoint and supports public clients (token_endpoint_auth_method: "none" plus PKCE S256) need no pre-provisioned OAuth app at all. At first connect the broker registers a client on the fly and stores it on the connection:

  • Registration request: client_name Paperclip (<instance host>), redirect_uris = the instance's own callback, grant_types ["authorization_code", "refresh_token"], response_types ["code"], token_endpoint_auth_method "none".
  • The issued client_id is persisted in the connection's OAuth config and any issued client_secret becomes a company_secrets ref. The registered client is reused for every later authorize/refresh on that connection — re-registering orphans prior grants on providers that bind grants to the client.
  • Env-registered clients always win: when PAPERCLIP_TOOL_OAUTH_<PROVIDER>_CLIENT_ID/_SECRET are configured, the broker uses them (customer ownership) and skips registration. List both customer and dcr in the method's ownershipModes when the vendor supports both.

Since PAP-17087, DCR is one of four registration tiers, and ownershipModes gates only the curated path. The broker resolves a client in this order: a deployment-preconfigured client, then a Client ID Metadata Document when the authorization server advertises one (requires a public HTTPS PAPERCLIP_PUBLIC_URL), then DCR, then client credentials the operator preregistered and pasted in. A URL-only connection with no AppDefinition may use the CIMD and DCR tiers too, but only after validated protected-resource and authorization-server discovery produced a metadata document. Registered client material is bound to the issuer, MCP resource URL, callback URI, and company; when a binding moves, a Paperclip-minted client re-registers and an operator-supplied one asks the operator to re-enter it. Full detail in Connecting any remote MCP server.

For a curated entry, ownershipModes still decides whether Paperclip may dynamically register on that vendor's behalf: omit dcr for a vendor that must not be auto-registered, and the broker will not fall through to the generic registration path for it.

DCR needs neither Paperclip ID nor Paperclip Connect. DCR is always instance-local (ratified in the PAP-14828 connector-service spec, section 10 item 8.4: "DCR is always instance-local; the service has no DCR involvement"). Each instance registers its own public client with the vendor and uses its own /api/tools/oauth/callback redirect. Cloud-hosted and self-hosted instances use the SAME path — the only per-instance difference is the hostname inside the redirect URI. id.paperclip.ing authenticates operators only and never holds resource tokens; connect.paperclip.ing is a fallback only for providers that genuinely require a pre-registered public redirect, which a DCR provider by definition does not.

Redirect-URI constraints

Vendors restrict what redirect_uris a dynamic client may register. Record the probed constraint in the AppDefinition redirectConstraints field and enforce it before starting OAuth. The first supported value is https-or-loopback-http (Notion's rule): HTTPS on any host — public or private — or plain HTTP only on loopback (localhost, *.localhost, ::1, 127.0.0.0/8). A plain-HTTP non-loopback origin fails fast with oauth_redirect_origin_unsupported ("This provider requires an HTTPS or loopback origin. Configure TLS before connecting.") and a pointer to the TLS deployment docs, instead of a confusing vendor-side invalid_redirect_uri. Probe the constraint with real registration attempts before writing the manifest — the redirect-URI rule and browser-reachability are independent axes; a private HTTPS host can be fine even when plain HTTP is not.

Documentation standards for every connection doc

Every connection doc — playbook appendix, proposal, or user-facing doc — must include all three of the following (they are part of the template below):

  1. Service involvement statement. Say explicitly whether Paperclip ID or Paperclip Connect participates in the flow. For RFC 7591 DCR providers the answer is always: neither — DCR is instance-local and cloud vs self-hosted use the same path.
  2. Sequence diagram + exact endpoints. A sequence diagram of how the connection works, and the exact paths/endpoints used for auth: authorize, token, registration (if DCR), and the Paperclip callback. Keep mermaid sources next to the doc; do not put semicolons inside mermaid message text (they parse as statement separators).
  3. Administrator setup instructions. Step-by-step: what (if anything) an admin must register — callback URLs? client credentials? nothing, for DCR? — where to register it, and how to verify the connection works end to end.
  4. Provider consent interaction guards. During browser smoke tests, do not classify a disabled consent button as an OAuth failure until the page has received a real pointer or keyboard interaction and any documented delay has elapsed. For example, Sentry intentionally enables its upstream Approve button one second after the first interaction. Record this separately from Paperclip callback, token-exchange, and MCP health failures.

Template

Copy this section into a connector proposal or implementation issue.

## Vendor

- App key:
- App name:
- Owner:
- First-30 classification: MCP-direct / OpenAPI-shim / vendor-deep-wrapper
- Reason for classification:
- Security tier: S1 / S2 / S3 / S4
- Plugin needed? No / Yes, because:

## Transport And Auth

- Transport:
- Endpoint or approved template:
- Auth mode: OAuth / API key / app-installation / none
- OAuth scopes or key scope:
- Credential owner: company / user-delegated / app-installation
- Secret storage: company_secrets refs only
- Revocation behavior:

## Connection Flow (mandatory)

- Sequence diagram: <mermaid source or rendered image  REQUIRED for every connection doc>
- Auth endpoints (exact paths):
  - Authorize:
  - Token:
  - Registration (if DCR):
  - Discovery (.well-known), if any:
  - Paperclip callback: `/api/tools/oauth/callback` (or n/a)
- Redirect constraints (probed): none / https-or-loopback-http / requires-public-redirect
- Paperclip ID / Paperclip Connect involvement: <"none — DCR is instance-local; cloud and self-hosted use the same path" for RFC 7591 providers; otherwise name the role>

## Administrator Setup (mandatory)

- What the admin must register (callback URLs? client credentials? nothing for DCR?):
- Where to register it:
- Instance prerequisites (TLS, base URL, feature flags):
- How to verify the connection works:

## Resource Filters

- Required filters:
- Optional filters:
- Write-enabling filters:
- Filters enforced by:

## Manifest

- schemaVersion: 1
- slug:
- name:
- description:
- categories:
- branding and provenance:
- docsUrl:
- method key and label:
- transport: mcp_remote / local_stdio / rest_api
- auth: oauth / api_key / none
- ownershipModes: dcr / customer / platform_shared / platform_provisioned
- grantKinds: organization / user
- oauthStrategy and connectorProfile, if reviewed:
- capabilityProfile and variants, if needed:
- defaults: endpoint/template/discovery/OAuth hints/scopes/tool defaults
- tenantFields and extensionFields:
- credentialFields:
- keyPlacement or credentialSources:
- configRequirements:
- guidanceMd, warnings, and consoleLinks:
- riskTier and requiredResourceFilters:
- urlPatterns:
- setupPrerequisite and redirectConstraints:
- availability:

## Actions

| Tool | Risk | Default status | Filters | Approval default | Audit fields | Negative case |
| --- | --- | --- | --- | --- | --- | --- |
| | read/write/destructive | active/quarantined/disabled | | allow/ask-first/block | | |

## Wizard Path

- User path:
- Configuration steps:
- Error states:
- Redacted metadata shown:

## Governance Defaults

- Default profile:
- Profile bindings:
- Policies:
- Quarantine rules:
- Rate limits:

## Validation Hook

- Real-vendor smoke issue:
- Connect evidence:
- Catalog evidence:
- Allowed read:
- Governed write (Allowed by default; operator policy may narrow it):
- Denied/quarantined case:
- Revoke:
- Audit:

Appendix: Linear Dry Run

This dry run applies the template to Linear, one of the PAP-2432 Batch A providers.

Vendor

  • App key: linear
  • App name: Linear
  • First-30 classification: MCP-direct with a thin GraphQL/resource-filter wrapper if the hosted MCP server cannot enforce all filters itself.
  • Reason for classification: Linear has a hosted MCP endpoint shape in the current gallery, and the first-30 matrix calls Linear a direct MCP/GraphQL thin-wrapper provider.
  • Security tier: S2, because it exposes product planning data and narrow issue mutations but not payments, tenant admin, or production infrastructure.
  • Plugin needed: No. The default gallery card, OAuth connect, resource filters, action catalog, profiles, policies, and audit cover the required UX. A plugin would only be warranted later for custom Linear dashboards or background sync workers.

Transport And Auth

  • Transport: mcp_remote
  • Endpoint: https://mcp.linear.app/mcp
  • Auth mode: OAuth
  • OAuth scopes: the reviewed read and write set. Linear is S2, so the current tier default allows reviewed writes; operators may still narrow them with profiles and policies.
  • Credential owner: company connection backed by user/workspace consent.
  • Secret storage: OAuth token material stored as company_secrets refs; no token in agent env, project env, comments, logs, or screenshots.
  • Revocation behavior: disabling or revoking the connection immediately removes Linear tools from agent sessions and denies brokered execution on the next gateway check.

Resource Filters

  • Required filters: workspace, team.
  • Optional filters: project, label, cycle, issue status.
  • Write-enabling filters: team plus project or label/cycle filter for create/update; comment-only writes may allow team-only with explicit policy.
  • Enforced by: gateway policy selectors, wrapper-side argument validation, and vendor request construction. UI filter pickers are convenience only, not the enforcement boundary.

Manifest Sketch

{
  "schemaVersion": 1,
  "slug": "linear",
  "name": "Linear",
  "description": "Create, update, and read Linear issues.",
  "categories": ["productivity"],
  "branding": { "logoUrl": "/brands/apps/linear.svg" },
  "urlPatterns": ["https://mcp.linear.app/*"],
  "methods": [
    {
      "key": "mcp-oauth",
      "transport": "mcp_remote",
      "auth": "oauth",
      "ownershipModes": ["customer"],
      "whenToUse": "Use the provider-hosted connection for the quickest setup.",
      "defaults": {
        "serverUrl": "https://mcp.linear.app/mcp",
        "authorizationEndpoint": "https://linear.app/oauth/authorize",
        "tokenEndpoint": "https://api.linear.app/oauth/token",
        "scopesHint": ["read", "write"]
      },
      "guidanceMd": "Register a Linear OAuth app and add Paperclip's redirect URI before connecting.",
      "riskTier": "S2",
      "requiredResourceFilters": ["workspace", "team", "project"]
    }
  ]
}

Actions

Tool Risk Default status Filters Approval default Audit fields Negative case
linear.search_issues read active after catalog review workspace, team, project, label, status allow when profile includes Linear reads query summary, team/project ids, result count Granted agent cannot search a disallowed team.
linear.get_issue read active after catalog review workspace, team, issue id allow when profile includes Linear reads issue id, team/project ids Ungranted agent cannot list or invoke the tool.
linear.create_issue write active workspace, team, project, label allow under S2 default team/project ids, title hash, created issue id Missing project/team filter denies.
linear.comment_issue write active workspace, team, issue id allow under S2 default issue id, comment body redaction summary Agent cannot comment on a disallowed issue.
linear.update_issue_status write active workspace, team, issue id, allowed statuses allow under S2 default issue id, old/new status if returned Revoked connection blocks retry.

No destructive Linear action should ship in the first pass. If one becomes part of the normal catalog, re-evaluate the method tier and changed-tool quarantine before accepting it as an S2 Allowed action.

Wizard Path

  1. Operator opens Apps and selects Linear.
  2. Operator clicks Connect and completes Linear OAuth.
  3. Paperclip stores OAuth material in company_secrets and shows redacted workspace/account metadata.
  4. Operator selects workspace/team/project filters and reviews the S2 Allowed action defaults.
  5. Paperclip runs health check and catalog refresh.
  6. Operator binds the Linear read profile to a company, project, agent, routine, or issue scope.
  7. Write actions are Allowed by the current S2 default unless the operator narrows them with profiles or policies.

Governance Defaults

  • Default profile: include the reviewed Linear actions for the selected scope.
  • Policy defaults: S2 actions are Allowed. Operators may narrow specific writes.
  • Quarantine: enable changed-tool quarantine before relying on it; the manifest declaration alone does not activate it.
  • Rate limits: apply a per-connection query/write budget to protect vendor quota and avoid noisy issue edits.
  • Audit: log connect, config/filter changes, grant changes, action requests, allowed/denied calls, revoke, and catalog quarantine events.

Validation Hook

Linear's real-vendor evidence belongs in PAP-12373. The smoke pass should prove:

  • OAuth connect succeeds with a customer-created Linear OAuth app (or an explicitly reviewed external credential source) and the instance callback URI.
  • Catalog discovery returns the expected Linear issue actions.
  • A read call against an allowed team succeeds.
  • linear.create_issue executes under the S2 Allowed default and remains bound by resource filters and current policy.
  • A call against a disallowed team/project is denied.
  • Revocation removes Linear tools and blocks execution.
  • Audit rows include company, connection, run/issue, agent/user actor, tool, decision, reason code, and outcome.

AppDefinition catalog authoring

Connector proposals now target the versioned AppDefinition contract in packages/shared/src/types/app-definition.ts. Seed data is one JSON file per provider under packages/shared/src/app-definitions/; regenerate Wave 1 with pnpm connections:ingest-app-definitions. The generator parses all 99 captured templates, validates required placeholders, OAuth ownership modes, and API-key placement, and produces deterministic output for review. FIRST-30 remains authoritative for riskTier and requiredResourceFilters; managed ownership modes stay data-visible but runtime-hidden until availability is injected.

Appendix: Notion Dry Run (MCP-Direct With DCR)

This dry run applies the template to Notion, the first MCP-direct connector to ship with RFC 7591 dynamic client registration (PAP-16637; server implementation PAP-16649, PR #11009). Unlike the Linear appendix, every endpoint and constraint below comes from a live request log, not vendor docs alone.

Vendor

  • App key: notion
  • App name: Notion
  • First-30 classification: MCP-direct. Notion ships an official hosted MCP server; its ~20 notion-* tools map directly to Paperclip grants.
  • Reason for classification: no shim or wrapper needed — the hosted server speaks Streamable HTTP, which server/src/services/mcp-http.ts already handles. The FIRST-30 matrix's "thin wrapper for block/database policy" is explicitly deferred; v1 enforcement is gateway policy plus filters-as-config.
  • Security tier: S3 — workspace content read/write, but no payments, tenant admin, or production infrastructure.
  • Plugin needed: No. Gallery card, OAuth connect, filters, catalog, profiles, policies, and audit cover the UX.

Transport And Auth

  • Transport: mcp_remote
  • Endpoint: https://mcp.notion.com/mcp (Streamable HTTP; /sse fallback exists)
  • Auth mode: OAuth, endpoints resolved by discovery (RFC 9728 → RFC 8414), public client via RFC 7591 DCR with PKCE S256 mandatory. Discovery runs because notion.json deliberately ships only serverUrl — no authorizationEndpoint/tokenEndpoint hints, which would otherwise take precedence and be used verbatim (see "MCP-Direct Connections" above).
  • Ownership modes: dcr (default, zero setup) and customer (env-registered classic integration via PAPERCLIP_TOOL_OAUTH_NOTION_CLIENT_ID/_SECRET, which always wins when set).
  • Token behavior: access tokens last ~8 h (expires_in authoritative). Refresh tokens rotate on every refresh — the old token is invalidated (at most 2 valid per grant) and replaying a stale one can revoke the whole grant, so the broker persists the rotated token before publishing the new access token and serializes refresh per connection. Absolute expiry 180 days, inactivity expiry 30 days. invalid_grant on refresh is terminal: clear tokens, require re-auth, never retry.
  • Secret storage: access/refresh tokens and any DCR client_secret are company_secrets refs; the DCR client_id persists on the connection and is reused — re-registering would orphan prior grants.
  • Revocation behavior: disabling or revoking the connection removes notion-* tools from agent sessions and denies brokered execution on the next gateway check.

Connection Flow (mandatory)

Paperclip ID / Paperclip Connect involvement: none — DCR is instance-local (PAP-14828 spec section 10 item 8.4); cloud-hosted and self-hosted use the same path. The only per-instance difference is the hostname in the redirect URI.

Auth endpoints (exact paths, from the live discovery chain):

Role Endpoint
MCP server https://mcp.notion.com/mcp
Protected-resource metadata (RFC 9728) https://mcp.notion.com/.well-known/oauth-protected-resource/mcp
AS metadata (RFC 8414) https://mcp.notion.com/.well-known/oauth-authorization-server
Authorize https://mcp.notion.com/authorize
Token (exchange + refresh) https://mcp.notion.com/token
Registration (RFC 7591 DCR) https://mcp.notion.com/register
Paperclip connect (wizard) POST /api/companies/:companyId/tools/apps/connect
Paperclip OAuth start POST /api/tools/oauth/:connectionId/start
Paperclip callback GET /api/tools/oauth/callback

Redirect constraints (probed): https-or-loopback-http.

sequenceDiagram
    autonumber
    actor U as User's browser
    participant UI as Paperclip UI<br/>/PAP/apps/connect?source=notion
    participant S as Paperclip instance server<br/>(cloud or self-hosted — same path)
    participant M as mcp.notion.com<br/>(MCP server + OAuth AS)
    participant N as Notion web<br/>(app.notion.com, notion.com)

    U->>UI: Click "Connect" (deep link ?source=notion)
    UI->>S: POST /companies/:id/tools/apps/connect { appKey: "notion" }
    S->>M: GET /.well-known/oauth-protected-resource (RFC 9728)
    M-->>S: authorization_servers → mcp.notion.com
    S->>M: GET /.well-known/oauth-authorization-server (RFC 8414)
    M-->>S: authorize / token / registration endpoints
    alt First connect on this instance (no stored client, no env client)
        S->>M: POST registration_endpoint (RFC 7591 DCR, public client, PKCE-only)
        M-->>S: client_id (persisted, REUSED for every later connect)
    else Client already known
        S->>S: Reuse stored DCR client_id (or env-registered client if configured)
    end
    S-->>UI: auth.startUrl (authorize URL + PKCE S256 challenge + state)
    UI->>U: Redirect browser to startUrl
    U->>M: GET /authorize?client_id + code_challenge + state
    M->>N: 302 to app.notion.com/install-integration
    N->>N: notion.com/login (only if signed out)
    N-->>U: Consent page: pick workspace, approve integration
    U->>S: 302 to GET /api/tools/oauth/callback?code&state (instance's OWN callback)
    S->>M: POST token_endpoint (code + code_verifier)
    M-->>S: access_token (~8 h) + rotating refresh_token
    S->>S: Store tokens as company_secrets refs (server-side only)
    S-->>U: Redirect to wizard actions/review step (?oauth=connected)
    Note over S,M: Later: agent runs reach notion-* tools via the managed MCP gateway.<br/>Server refreshes ahead of use — each refresh ROTATES the refresh token.

Dry-Run Request Log (PAP-16649, 2026-08-06/07)

The verified request sequence for a first connect:

  1. GET https://mcp.notion.com/mcp401 with WWW-Authenticate naming https://mcp.notion.com/.well-known/oauth-protected-resource/mcp.
  2. GET https://mcp.notion.com/.well-known/oauth-protected-resource/mcp200; authorization server https://mcp.notion.com, scope default.
  3. GET https://mcp.notion.com/.well-known/oauth-authorization-server200; /authorize, /token, /register; token_endpoint_auth_method none supported; PKCE S256 supported.
  4. POST https://mcp.notion.com/register (RFC 7591).
  5. Browser GET https://mcp.notion.com/authorize → Notion consent (app.notion.com/install-integration, notion.com/login if signed out).
  6. POST https://mcp.notion.com/token for code exchange and every refresh.
  7. POST https://mcp.notion.com/mcp for MCP traffic.

Redirect-URI probes against /register:

Probed redirect_uris value Result
http://paperclip-dev:3100/api/tools/oauth/callback 400 invalid_redirect_uri — "Redirect URI must use HTTPS unless it is a loopback HTTP URI"
https://paperclip-dev:3100/api/tools/oauth/callback Accepted — private host is fine over HTTPS
http://localhost:3100/api/tools/oauth/callback Accepted
http://127.0.0.1:3100/api/tools/oauth/callback Accepted

Hence redirectConstraints: "https-or-loopback-http" in notion.json, and the broker's fail-fast oauth_redirect_origin_unsupported error for plain-HTTP non-loopback origins.

Administrator Setup (mandatory)

  • What the admin must register: nothing. Notion's authorization server supports RFC 7591 DCR, so the instance registers its own public client on first connect. No Notion integration, no client credentials, no callback registration, no Paperclip ID or Paperclip Connect involvement.
  • Optional escape hatch: to use a pre-registered classic Notion integration instead, set PAPERCLIP_TOOL_OAUTH_NOTION_CLIENT_ID and PAPERCLIP_TOOL_OAUTH_NOTION_CLIENT_SECRET; the env client always takes precedence (customer ownership).
  • Instance prerequisites: the instance base URL must be HTTPS on any host or loopback HTTP (Notion's redirect-URI rule). A plain-HTTP non-loopback origin gets "This provider requires an HTTPS or loopback origin. Configure TLS before connecting." — add TLS first (e.g. a tailscale cert, as paperclip-dev did). Apps is a standard product surface and /apps/* routes are always available. The connecting user must be allowed to install integrations in their Notion workspace.
  • How to verify: visit /PAP/apps/connect?source=notion, complete the Notion consent flow, and land on the wizard's actions step listing notion-* tools. Then confirm an agent run sees Notion tools through the runtime MCP gateway and that a write call such as notion-create-pages follows the current S3 Allowed policy. Add a narrower Ask-first rule separately when the company wants approval for that action.

Resource Filters

  • Required filters: workspace, page, database (per FIRST-30).
  • Optional filters: object type, database/data-source scope.
  • Write-enabling filters: workspace plus page/database scope for create/update.
  • Enforced by: gateway policy plus filters-as-config in v1; the FIRST-30 "thin wrapper for block/database policy" is explicitly deferred. Notion-side scoping also applies — the consent step lets the user share only selected pages/databases with the integration.

Manifest Sketch

The shipped packages/shared/src/app-definitions/notion.json (regenerate via pnpm connections:ingest-app-definitions):

{
  "schemaVersion": 1,
  "slug": "notion",
  "name": "Notion",
  "description": "Read and update pages in your Notion workspace.",
  "urlPatterns": ["https://mcp.notion.com/*"],
  "methods": [
    {
      "key": "mcp-oauth",
      "transport": "mcp_remote",
      "auth": "oauth",
      "ownershipModes": ["customer", "dcr"],
      "defaults": { "serverUrl": "https://mcp.notion.com/mcp" },
      "riskTier": "S3",
      "requiredResourceFilters": ["workspace", "page", "database"]
    }
  ],
  "redirectConstraints": "https-or-loopback-http"
}

Actions

Notion's hosted server exposes ~20 notion-* tools. Representative risk classes below; the full catalog review with per-tool defaults is PAP-16652 (P4). Changed-action quarantine applies only when the connection explicitly enables quarantineNewEntries.

Tool Risk Default status Filters Approval default Audit fields Negative case
notion-search read active after catalog review; plan-gated by Notion (needs Notion AI) — may list but fail at call time workspace allow when profile includes Notion reads query summary, result count Ungranted agent cannot invoke.
notion-fetch read active after catalog review workspace, page, database allow when profile includes Notion reads page/database id Fetch outside shared pages fails Notion-side and is audited.
notion-create-pages write active after catalog review workspace, page, database allow under S3 default parent id, title hash, created page id Missing workspace/page filter denies.
notion-update-page write active after catalog review workspace, page allow under S3 default page id, redaction summary Revoked connection blocks retry.
notion-query-data-sources read active after catalog review workspace, database allow when profile includes Notion reads data-source id, result count Granted agent cannot query a disallowed database.

No destructive Notion action ships in this worked example. A future delete/archive/bulk action needs explicit risk review; normally classify the method S4 or add a reviewed narrow policy and tests before enabling it.

Wizard Path

  1. Operator opens /PAP/apps/connect?source=notion (or the Notion gallery card → Connect). The deep link POSTs connect immediately and redirects the browser to auth.startUrl.
  2. Operator completes Notion consent (workspace picker → approve).
  3. Notion redirects to the instance's own GET /api/tools/oauth/callback; Paperclip exchanges the code, stores token material in company_secrets, and returns the operator to the wizard (?oauth=connected).
  4. Operator confirms resource filters and reviews the S3 Allowed action defaults.
  5. Paperclip runs health check and catalog refresh; notion-* tools appear on the actions step.
  6. Reviewed write actions are Allowed by the current S3 default unless the operator narrows them with profiles or an Ask-first policy.

Error state: on a plain-HTTP non-loopback instance, step 1 fails fast with the TLS guidance error above — the operator never reaches Notion.

Governance Defaults

  • Default profile: reviewed Notion actions for the selected resource scope.
  • Policy defaults: S3 actions are Allowed. Operators may narrow page creation, updates, or comments with profiles or Ask-first rules.
  • Quarantine: enable quarantineNewEntries before relying on changed-tool quarantine; the manifest declaration alone does not activate it.
  • Rate limits: per-connection search/fetch budget to protect vendor quota.
  • Audit: log connect, DCR registration, config/filter changes, grant changes, action requests, allowed/denied calls, token refresh failures, revoke, and catalog quarantine events.

Validation Hook

End-to-end evidence belongs to PAP-16654 (P6) and the PAP-12373 matrix:

  • Zero-setup OAuth connect succeeds on https://paperclip-dev.tail29c1aa.ts.net/PAP/apps/connect?source=notion with no pre-provisioned OAuth env vars (proves DCR).
  • Catalog discovery lists the expected notion-* tools and applies the connection's declared changed-tool behavior.
  • An agent run sees Notion tools through the managed runtime MCP gateway.
  • notion-create-pages executes under the S3 Allowed default while respecting resource filters and any narrower company policy.
  • Revocation removes Notion tools and blocks execution.
  • Audit rows prove actor, run/issue context, connection, tool, decision, reason code, and outcome.