Files
PaperClipAI/doc/connections
DottaandClaude Opus 5.5 944b1c64d7 feat(mcp): keep long provider calls durable with the MCP Tasks extension (#21263)
## Thinking Path

> - Paperclip is the open source app people use to manage AI agents for
work
> - Agents use connected apps through Paperclip's MCP client in the tool
gateway
> - Some provider actions run for minutes: CI runs, deploys, batch jobs
> - Today such a call is lost if it outlives the tool timeout or if
Paperclip restarts or deploys mid-call, and a retry can start the work a
second time
> - Recovery must just work, and the target rate of runs that need a
manual retry is 0%
> - This pull request implements the MCP Tasks extension
(`io.modelcontextprotocol/tasks`, 2026-07-28) in the MCP client
> - The benefit is that a long provider call becomes a durable task: it
survives timeouts, restarts, and deploys, and it is never started twice

## Linked Issues or Issue Description

**Problem**
A long provider tool call holds an HTTP request open. When it passes the
tool timeout, or when Paperclip restarts or deploys during the call, the
result is lost. The agent then often repeats the call, and the provider
does the work again.

**Proposed solution**
Implement the client side of the MCP Tasks extension. Declare it on
stateless tool calls. Store each returned task durably, wait on it
during the call, keep polling it in the background after the call ends,
and let the same call attach to the stored task instead of sending
`tools/call` again.

**Why it matters**
CI, deploy, and batch actions are where agents lose the most work today.
A durable task handle makes these actions safe across timeouts,
restarts, and deploys, without a person.

Refs #21119, Refs #21140, Refs #21219 (the MCP Apps, stateless MCP
client, and MCP Skills work this builds on).

## What Changed

- **Off by default:** MCP Tasks are behind the new instance experimental
setting `enableMcpTasks` ("Durable app actions (MCP Tasks)"), default
off. When it is off, Paperclip never declares the extension, nothing
attaches to or polls tasks, pending rows do not block idle sleep, and
behavior matches master. A test covers the off state.

- `mcp-http.ts`: the tasks extension is declared per request, only on
HTTP `tools/call` and `tasks/*`. Discovery, listings, and stdio servers
do not declare it. `tasks/*` requests put the task ID in `Mcp-Name`.
- New `mcp-tasks.ts`: parses task results, bounds the server's timing
hints, sends `tasks/get`, `tasks/update`, and `tasks/cancel`, and seals
task IDs at rest (AES-256-GCM, bound to company and connection; a keyed
hash gives uniqueness).
- New company-scoped table `tool_mcp_tasks` (migration `0475`,
online-compatible). It links the task to the call, run, task,
connection, and grant. It stores no provider result.
- `tool-gateway.ts`:
- A returned task is recorded, then polled up to the normal tool
timeout. Its result goes through the same validation, redaction, and
recording as a direct result.
- Past the timeout, or when the caller disconnects, the agent gets
`provider_task_pending` ("call again with the same arguments"). The task
stays durable.
- A call with the same endpoint, tool, full arguments, grant, agent,
task, and actor attaches to the undelivered task. A retry with the same
idempotency key also attaches, and it can never start different work
under that key.
- Input: URL requests show the provider link (Paperclip never answers
them). Form requests become a question card; only the person's answer is
sent with `tasks/update`. Other requests stop with a clear "needs input"
error.
- Stopping the run sends `tasks/cancel` and marks the task cancelled. A
caller that only disconnects keeps the task.
- `failed` becomes a tool error with the provider's message. An unknown
task or a passed TTL becomes `provider_task_expired`. Every task stops
being tracked at its TTL, or after 7 days at most.
- Audit entries: `tool_gateway.task_started`, `task_input_sent`,
`task_cancelled`.
- New `mcp-task-work.ts`, registered at startup with the shared deadline
scheduler: it resumes polling pending tasks after a restart, cancels
tasks whose task closed or run was stopped, and wakes the assignee once
when a task settles after its run ended.
- Docs: `doc/connections/MCP-TASKS.md`, linked from
`doc/connections/README.md`.

Background work (per `rules/sleep-aware-background-work.mdc`): trigger
is a task row write or release (`mcp-tasks` queue) and one startup pass.
The deadline is the earliest `next_poll_at`. Cadence is the server's
`pollIntervalMs` within 5 s – 5 min, with backoff; 5 min while waiting
for a person. It is quiet when no row has a next poll. This is the
narrow external-source exception: providers give no wake Paperclip can
receive, and polling stops with the task.

## Verification

- `server/src/__tests__/mcp-tasks.test.ts`: 20 tests with a fake
2026-07-28 server that returns tasks only to a client that declares the
extension and checks `Mcp-Name`. They cover:
- polling to completion, with the result identical to a synchronous
call;
  - timeout, then an attaching call with no second `tools/call`;
  - a keyed retry, and a reused key with different arguments (refused);
- restart mid-task: a fresh gateway with no in-memory state polls the
row, wakes the agent once, and the recovered run gets the result once
with no second `tools/call`;
  - no wake while the run is still going;
- URL input (link, never answered), and form input (question card, then
`tasks/update` with the answer);
- cancellation by the run, by closing the task, and no cancellation when
only the caller leaves;
- failure, an unknown task, and TTL expiry (ends in under 3 s, no hang);
  - another agent never attaches;
- synchronous calls unchanged and older servers never see the extension.
- Red/green: with attach disabled, the timeout, keyed-retry, and restart
tests fail (3). With `tasks/cancel` disabled, both cancellation tests
fail (2). Before `tools/call` declared the extension, 13 end-to-end
tests failed. All pass with the change.
- Existing suites pass: `mcp-protocol-versions`, `mcp-http`, `mcp-apps`,
`mcp-skills`, `server-startup-feedback-export`, `tool-gateway`,
`tool-gateway-discovery-http`, `remote-mcp-connectors` (342 tests).
- `tsc --noEmit -p server` is clean. `pnpm --filter @paperclipai/db
check:migrations` passes (the migration is classified
online-compatible). Snapshots pruned to five.
- Agent boot and time-to-first-token: no change in the boot or
first-turn path. The only new per-call cost is one indexed lookup on
stateless tool calls.
- Not run: a live provider that returns tasks. No public MCP server that
returns tasks was available without new credentials.

## Risks

- Shipped off by default. Review findings that only affect the
setting-on path after this merge are tracked as follow-ups.

- Migration: a new table with indexes and foreign keys to `companies`,
`tool_invocations`, and `tool_connections`. Online-compatible.
- Behaviour change only for servers that return tasks, which they may do
only to clients that declare the extension. Older servers and stdio
servers are unchanged.
- A task that settles while nobody is waiting stays attachable for 1
hour, so a deliberate repeat of the same call within that hour collects
the earlier result instead of starting new work.
- The wake after a background task settles uses the generic task wake;
the agent must call the tool again to collect the result.
- Follow-ups: `notifications/tasks` via `subscriptions/listen`; sampling
and roots input requests; showing a long action's progress in the task
chat.

> For core feature work, check [`ROADMAP.md`](ROADMAP.md) first and
discuss it in `#dev` before opening the PR. Feature PRs that overlap
with planned core work may need to be redirected — check the roadmap
first. See `CONTRIBUTING.md`.

## Model Used

- Claude Opus 5.5 (`claude-opus-5-5`), Anthropic, in Claude Code with
tool use (shell, file edits, local test runs).

## Checklist

- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have searched GitHub for duplicate or related PRs and linked
them above
- [x] I have either (a) linked existing issues with `Fixes: #` / `Closes
#` / `Refs #` OR (b) described the issue in-PR following the relevant
issue template
- [x] I have not referenced internal/instance-local Paperclip issues or
links (only public GitHub `#NNN` / `github.com/paperclipai/paperclip`
URLs)
- [x] My branch name describes the change (e.g. `docs/...`, `fix/...`)
and contains no internal Paperclip ticket id or instance-derived details
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [ ] All Paperclip CI gates are green
- [ ] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
- [ ] I will address all Greptile and reviewer comments before
requesting merge

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-13 08:32:06 -06:00
..

Apps, Connections, and Integrations

Audience: internal engineers and product contributors working on integrations.

Start here when adding a provider: Connection authoring runbook. It is the canonical agent tutorial from provider research and protocol classification through manifest generation, branding, secrets, deterministic tests, real-account proof, and PR submission.

Chat and email channels: conversation-channel contract, question inventory and quality bar, and per-provider question coverage index.

Runtime authentication: AI Connections.

MCP protocol versions (2026-07-28 stateless MCP and older handshake servers): MCP-PROTOCOL-VERSIONS.md.

MCP extensions: MCP Apps (interactive views in the task chat), MCP Skills (skills that connected apps publish, reviewed by the board and delivered to agents), and MCP Tasks (long provider calls that survive timeouts, restarts, and deploys).

Long-term memory: Experimental memory connectors.

Provider notes: Google Workspace, Gmail, Asana, Sentry, PostHog, Braintrust, Neon, Superagent, Gauge, Ditto, Circleback, 1Password, Unikraft, AgentMail, Telem.AI, iMessage Photon, Enterpret, Supabase, and Vercel. Optional credential custody: Vercel Connect.

Post-read action: classify a new integration request, pick the right Paperclip layer to change, and avoid creating a parallel connection framework.

Decision Record

Board decisions from PAP-13211 make the Apps v2 substrate on the PAP-10341 branch canonical:

  • D1: Apps v2 is the substrate. The active model is tool_applications, tool_connections, catalog entries, profiles, policy rules, action requests, gateway sessions, audit events, and runtime slots. Connections v1 is retired as an implementation path.
  • D2: one credential authority per connection, brokered projections. The default is the Paperclip instance vault: durable third-party credentials live in company_secrets as secret refs. A reviewed remote MCP method may instead opt in to Vercel Connect, in which case Vercel is the durable credential authority and Paperclip stores only the connector reference and redacted grant metadata. A connection must never mix those two sources. Adapter config, plugin config, harness credential files, and run environments may receive only brokered or projected credentials.
  • D3: the vocabulary and three-door IA are product law. The default product doors are Apps, Connections, and Review. Protocol and operator-depth concepts live behind Developer or Advanced surfaces.
  • D4: unification lands on PAP-10341. Pages, CircleBack-style harness MCP OAuth, provider gallery work, and plugin-provided integrations converge on this branch instead of spawning new integration substrates.
  • D5: inbound stays thin. External clients that call Paperclip use scoped Paperclip tokens and existing profiles/rules. They do not get a separate permission model.

Canonical Object Model

Use connection as the unifying noun. A connection is four things:

  1. A stored credential reference.
  2. A capability catalog.
  3. A governance layer.
  4. An audit trail.

Everything else is an axis on that object:

Axis Values It answers
Direction outbound, inbound Who is the client?
Transport MCP, native REST/OpenAPI, OAuth app install, webhook How do bytes move?
Auth mode OAuth, API key/PAT, app installation, none What does the secret represent?
Credential owner company, user, run Whose identity acts?
Packaging catalog entry, plugin, skill How does it ship?

MCP is a transport, not a product category. "Install the Discord app", "connect Google Drive", and "add an MCP endpoint" all produce governed connections with different transport/auth values.

Layer Stack

When you are unsure where a change belongs, place it on the narrowest layer that solves the problem:

Layer Owns Examples
Surface user-facing Apps, Connections, Review, Developer/Advanced screens gallery cards, setup wizard, review queue
Governance profiles, bindings, allow/ask-first/block rules, quarantine, audit read-only profile, ask-first write policy
Capability action catalogs, schemas, risk classes, changed-tool review search_issues, create_comment, schema hash
Credential company_secrets, OAuth broker, credential resolver, token broker Slack bot token ref, Google OAuth refresh token ref
Identity actor attribution and token exchange board user, agent run, first-party service identity
Transport how the external system is reached remote HTTP MCP, local stdio, REST/OpenAPI, webhook

The agent should not hold a durable provider credential. It should hold a Paperclip run/session token; the server or broker resolves the connection, checks governance, invokes the provider, and writes audit.

Identity vs. connections

Signing a user in and connecting a resource are different planes with different owners, different token profiles, and different homes. Do not merge them. This section is the public, connections-side statement of the identity model so connector implementers inherit the rule without depending on private identity-service documentation or re-deriving it.

Plane Question Lives where Token profile
P1. Sign-in methods Who are you? paperclip-id (id.paperclip.ing → Account) Minimal-scope provider tokens (openid email profile), used once to authenticate, encrypted at rest, never exported
P2. Connections (Apps) What may your agents touch? Paperclip App instances (tool_connections), acquired via the connect broker for hosted + self-hosted Rich-scope, long-lived resource tokens in the instance's encrypted vault; per-agent grants; risk-tier policy defaults
P3. Login with Paperclip Who may authenticate against us? paperclip-id OIDC provider + DB-backed client registry Our ES256 ID/access tokens issued by us to registered RPs (instances, the broker, future third parties)

Everything in doc/connections/ — the First-30 matrix, the connection authoring runbook, and the connect-broker work — lives on plane P2. It never acquires, stores, or brokers a P1 sign-in token.

The standing rule (D7)

Adopted as a standing rule (decision D7) with the identity-model plan. State it verbatim in any P2 design so the app-store work cannot drift into merging the planes:

Sign-in tokens are never reused as resource tokens; id.paperclip.ing never stores resource tokens; no connections hub on the ID service.

P2 tokens flow broker → instance vault as pass-through only; the id.paperclip.ing Account page therefore must not grow a "Connections" hub. The reasons to hold the planes apart (from the plan §3):

  • Scope discipline. Sign-in wants the narrowest grant; connections want deliberately broad ones. One button that does both is how you grant repo access just to log in.
  • Blast radius. id.paperclip.ing holding every customer's Vercel/Slack/GitHub resource tokens would make it the single juiciest target in the fleet; the broker is intentionally pass-through.
  • Self-hosted symmetry. Instances own their vaults, so self-hosters don't depend on our uptime to use their own connections.
  • Legibility. Sign-in and connections answer different user questions, and every product we benchmarked (Vercel, Railway, GitHub, Google) keeps them on separate pages with separate names.

The explicit Vercel Connect exception does not change D7 or merge P1 and P2. The operator chooses Vercel as the P2 credential authority for an individual connection. id.paperclip.ing is not involved, and neither sign-in tokens nor provider tokens pass through it. The deployment's Vercel access token or workload OIDC identity is bootstrap authority for that external vault, not a provider resource credential.

Naming alignment

Use the surface-correct name for each plane; they intentionally differ:

Surface Plane Name to use
Paperclip App instances P2 "Connections"
id.paperclip.ing Account P1 "Ways to sign in"
id.paperclip.ing admin P3 "OIDC clients" (until the app store productizes it)

Packaging Rule

Default to a catalog entry when an integration can be described as metadata: manifest, auth config, action catalog, resource filters, and policy defaults.

Use a plugin only when the integration needs product code such as custom UI pages, its own tables, workers, migrations, routines, or specialized ingestion. A plugin may bundle catalog entries, but it must not bypass the connection, profile, policy, credential, and audit model.

Use a skill for agent instructions. Skills may use connections; they must not own durable tokens.

Canonical Docs

  • Glossary defines product and internal terms.
  • Identity vs. connections is the public statement of the P1/P2/P3 boundary and the D7 standing rule for connections work.
  • Security threat model harvests the keeper from PAP-2359 and maps it onto Apps v2.
  • First-30 matrix harvests the keeper from PAP-2432 and is the source matrix for connector playbook work.
  • Connecting any remote MCP server is the baseline: how an operator connects a standards-compliant remote MCP endpoint with no Paperclip code change, and how sign-in resolves a client.
  • Connection authoring runbook is the one end-to-end, agent-executable guide for adding a vendor as a catalog entry on Apps v2: research, connection-type selection, OAuth/API-key/generated-URL setup, encrypted credential handling, branding, implementation, browser and live-provider testing, verification, and PR submission.
  • Vercel Connect operator guide documents the optional external credential source, deployment flags, runtime resolution, recovery, and smoke requirements.
  • MCP access governance remains the operator runbook for the current gateway, profile, policy, approval, runtime, and audit APIs.

Migration Notes

Connections v1 contributed useful policy, UX, and rollout thinking, but its implementation branch is no longer the target. When you see old tickets or code using connections, connection_grants, or a provider-directory mental model, translate the intent into Apps v2:

Connections v1 intent Apps v2 home
Provider directory Apps gallery / tool_applications
Configured provider instance Connection / tool_connections
Grant allowlist Profiles, profile bindings, policies
Resource filters Policy/profile conditions plus provider config
Tool broker Tool gateway and runtime supervisor
Connection UX tail Apps, Connections, Review, Developer/Advanced IA

Do not add new work to the retired v1 branch. If an old ticket still describes a valid product gap, retarget it to an active Apps v2 issue or close it as superseded with a link to the replacement.

For Slack chat bot creation, installation, recovery, and live qualification, see Automatic Slack app setup.