## Summary Release docs for Paperclip **v2026.1012.0** (parent tag commit `bba6a45f`). - **Window:** `v2026.1005.0` → `v2026.1012.0`. That is 132 commits and 1,242 files, with no compare truncation. - **Branch:** cut from `nightly` (`fe556e1`). Nightly had drafted against parent master up to `abbd8800`. The release was cut from master at `22a3ea3` (Oct 6), so nightly's last 39 commits are post-tag. - **Review:** every changed page was checked by hand against a checkout of the tag. The tag-accuracy gate was then re-run on the result. **Headline doc changes:** - **New pages:** - [Assistant connections (MCP)](docs/experimental/assistant-connections.md), with the [Assistant MCP API](docs/reference/api/assistant-mcp.md) and [`paperclipai mcp`](docs/reference/cli/mcp.md) references - [Saved connection instructions](docs/connectors/connection-instructions.md) - [Check AI account usage](docs/connectors/ai-usage.md) - [Task-pinned AI routing](docs/experimental/task-pinned-ai-routing.md) - [Neon](docs/connectors/neon.md) - **Breaking changes documented:** - Restore locks moved to SQLite (#14869). [Update Paperclip](docs/how-to/update-paperclip.md) now has a "stop everything first" upgrade and rollback flow. - Stored tool grants are enforced as written (#14915). This is covered in the [Tool Gateway API](docs/reference/api/tool-gateway.md) and the [Access model](docs/connectors/access-model.md). - **Other release coverage:** - The connection-instructions contract for HTTP and process adapters (`connectionInstructions`, `PAPERCLIP_CONNECTION_INSTRUCTIONS_FILE`). - Much shorter default instructions, and default config grants for new agents. - The `update_skill` tool. - Aggregator accounts gathered into one catalog. - Chat-first navigation and the combined inbox. - The rebuilt New task dialog. - "Model at capacity" and "Model unavailable". - Sonnet 5.5, GPT-6.1 Sol and new harness pins. - The supported remote Codex range, and Codex resuming after a server restart. - Watchdog thresholds of 5 and 15 minutes. - `tool_discovery_busy`. - The Docker provider pack. - A changelog entry, and 75 pages stamped `paperclip_version: v2026.1012.0`. ### ⚠ Post-tag leaks removed These document parent `master` features that are **not in v2026.1012.0**. They stay on `nightly` for the next release. - **Pages removed (7):** - `administration/agent-commentary.md` (b508a05c) - `guides/org/agent-identity.md` (a590ab76) - `reference/cli/runtime.md` (a6306ba6) - `connectors/custom-model-providers.md` (9f7057e1) - `connectors/enterpret.md` (88ff98b8) - `connectors/superagent.md` (8cbd21b3) - `connectors/telem.md` (f77fcbf4) Their nav entries and catalog rows are already removed, and no page links to them. - **Reverted to `main`:** - `roles-and-permissions.md`, `administration/company.md` and `add-a-human-teammate.md`. Operator's 14 default keys come from 582911ba. - The costs guide and `api/costs.md`. Durable accounting, by-user costs and per-account quota come from 799e4d55, eab93fd4 and 892b0b36. - `adapters/cursor-local.md` (a6306ba6) - `deploy/database.md` and `deploy/environment-variables.md`, which carried agent identity and complaints. - `browser-use-cloud.md` - `back-up-and-restore-a-company.md` - **Sections trimmed:** - The assistant MCP pages now match the tag: `paperclip:read`/`paperclip:write`/`offline_access` only, 10 direct tools, and no file transfer (2d0c1381, 99a9de99). - Server-host sign-in for Anthropic, OpenAI and xAI is back to the terminal-command flow (e38d6d16, 2ca0d26a). - OpenRouter is OpenCode-only again (9f7057e1). - Requested vs granted OAuth scopes are removed (ea8e6861). - PR review waits are removed (228f0e28). - Historical-question reminders and "No project" in the composer are removed (faa8e452, ee3c728b). - Plugin lifecycle hooks are removed (1cebdd4c). - Cursor on the Runner is removed (a6306ba6). - The aggregator "Waiting for sign-in" status is removed. - Source links pinned to post-tag commits now point at the tag commit. ### ⚠ Drift Checked against the tag: - **High:** the `operator` default grants were **real** drift. Fixed by reverting the post-tag tables. - **High:** the agent-identity paths and env vars were real at nightly but are post-tag. They're removed (see above). - **High:** `PAPERCLIP_WORKSPACE_GIT_SCAN_*`, `PAPERCLIP_WORKSPACE_GIT_SNAPSHOT_TIMEOUT_MS`, `PAPERCLIP_RUNNER_API_*` and `PAPERCLIP_CLOUD_CONNECTOR_*` are known false positives. They are read through helpers the checker doesn't scan. - **High:** `PAPERCLIP_ID_CONNECTOR_*` is documented as retired, which is intentional. - Verify: the `/accounting/*`, `costs/by-user` and `agents/{id}/identity` routes are no longer documented, because they're post-tag. - Verify: `agent-avatars`, `companies/import/transfers*` and `chats/{agentRef}` are known param/constant-registration false positives. ### ⚠ Verification Failures / review notes - **`reference/cli/mcp.md`:** the gate flags `--device` and `--url`. This is a false positive: both are `requiredOption`s in `cli/src/commands/mcp.ts` at the tag. - **New pages the gate can't auto-confirm:** each was verified by hand at the tag. They are assistant-connections, task-pinned-ai-routing, ai-usage and connection-instructions. - **`neon.md`:** says "open **Connectors**", following the Supabase guide. The parent's notes call the screen "Apps → Browse". - **`connection-intents.md`:** the AgentMail card details ("shares the key with every person in the company", "resumes after reload") are only partly confirmed. - **`task-chat.md`:** the queued-message details came from an earlier nightly and couldn't be traced to a specific commit. - **Upstream inconsistency:** the Codex adapter's in-code doc string (`codex-local/src/index.ts:181`) contradicts the code about the bypass flag. Our docs follow the code. ### Screenshots **126 of 350 screenshots are stale** across 26 routes (see `SCREENSHOTS_PENDING.md`). None were recaptured in this PR. ### Other open PRs - **#143** (Browser Use Cloud + Neon, targets `main`): both are now covered here. Suggest closing it. - **#140** (Enterpret, targets `main`): Enterpret is post-tag (88ff98b8). Suggest retargeting it to `nightly` or closing it. ## Checks - [x] `npm run docs:build` - [x] `npm run docs:test` (all suites) - [x] `sync:lint-links` and `sync:verify-nav`. The only orphans are the 7 pages above plus 4 intentional ones. - [x] `check-tag-accuracy` re-run. The only remaining flags are the ones explained above. - [x] Removed the 7 post-tag pages listed above (`cbece5b`), plus fixes from a second, independent verification pass - [ ] After the squash-merge, run `node scripts/sync/realign-nightly.mjs release/v2026.1012.0 --push`, then flip `nightly`'s `.sync-state.json` back to `branch_mode: nightly`. - [ ] Confirm the `docs/v2026.1012.0` tag and that docs.paperclip.ing serves the new changelog entry. 🤖 Generated with [Claude Code](https://claude.com/claude-code) --------- Co-authored-by: Claude Opus 4.8 (1M context) <[email protected]> Co-authored-by: DevRel Gemini <[email protected]> Co-authored-by: Paperclip <[email protected]> Co-authored-by: Dotta <[email protected]> Co-authored-by: ClaudeCoder <[email protected]>
7.5 KiB
paperclip_version, seo_title, seo_description
| paperclip_version | seo_title | seo_description |
|---|---|---|
| v2026.1012.0 | API Overview | The shared reference for Paperclip's JSON control-plane API — companies, agents, issues, approvals, costs, routines, secrets, activity, and dashboard state. |
API Overview
Paperclip exposes a JSON API for company control-plane work: companies, agents, issues, approvals, costs, routines, secrets, activity, and dashboard state. This page is the shared reference for how the API is structured before you jump into a specific resource.
Base URL
The API prefix is always /api.
In local development that usually means:
http://localhost:3100/api
Every endpoint in this section is relative to that prefix.
Authentication
Paperclip supports different request identities depending on the deployment mode and caller type:
| Caller | How it authenticates | Notes |
|---|---|---|
Board user in local_trusted |
Implicit board access | No login friction in local trusted mode. |
Board user in authenticated mode |
Session cookie from the web app | The UI uses the browser session; API clients can use the same session if they send cookies. |
| Board user with an API token | Authorization: Bearer <board-token> |
Board API keys are supported by the auth middleware. |
| Agent | Authorization: Bearer <agent-key-or-jwt> |
Agents use long-lived API keys or short-lived run JWTs. |
Practical rules:
- Board requests can operate across the companies the board user is allowed to access.
- Agent requests are always company-bound to the agent that owns the token or JWT.
- If you are calling the API from the browser UI, the session cookie is usually enough in authenticated mode.
- If you are calling the API from a script, use
Authorization: Bearer ....
Note: The server also reads
X-Paperclip-Run-Idon mutating requests during agent runs. That is mostly relevant for issue comments, checkout, and other run-linked actions.
External assistants can also connect as a human through the experimental Assistant MCP API, using user OAuth and explicit read and write scopes.
curl "http://localhost:3100/api/health"
const res = await fetch("http://localhost:3100/api/health");
const health = await res.json();
import requests
health = requests.get("http://localhost:3100/api/health").json()
Company Scoping
Most control-plane endpoints are company-scoped. That means the company ID must be present in the path, usually as:
/api/companies/{companyId}/...
Rules to keep in mind:
- Board users can only access companies they are actually allowed to see, unless they are implicit local board users or instance admins.
- Agent tokens can only access the company that issued the token.
- If you hit a company-scoped route with the wrong company, expect
403 Forbiddenor404 Not Founddepending on whether the server can tell you are unauthorized or the entity does not exist.
When a route is not company-scoped, it is usually because it operates on a global identity like a specific agent, issue, approval, or health check.
Request Format
Most API calls use JSON request bodies:
Content-Type: application/json
Common conventions:
- Use JSON for request bodies unless the route explicitly documents multipart upload or another format.
- List endpoints often return arrays or paginated collections depending on the resource.
- Validation is performed server-side with Zod schemas, so malformed payloads usually come back as
400.
Response Format
Successful responses return JSON.
Error responses also return JSON and usually look like:
{
"error": "Human-readable error message"
}
Some validation and domain errors also include a details field with structured context.
Error Codes
These are the most common status codes you will see:
| Code | Meaning | Typical cause |
|---|---|---|
400 |
Bad request | Validation failed, required fields were missing, or the payload shape was wrong. |
401 |
Unauthorized | No valid caller identity was provided. |
403 |
Forbidden | You are authenticated, but not allowed to perform the action. |
404 |
Not found | The resource does not exist, or it is outside your company scope. |
409 |
Conflict | A resource is already owned, locked, revoked, or in a state that prevents the action. |
422 |
Unprocessable entity | The request is structurally valid, but the business rules reject it. |
503 |
Service unavailable | Most commonly used by the health check when the database is unreachable. |
500 |
Internal server error | The server hit an unexpected failure. |
Two useful distinctions:
401means the server does not accept the caller identity.403means the caller is known, but not allowed to do this.
Health Check
GET /api/health
The health endpoint is the fastest way to check whether the server is up and whether the database is reachable.
It also reports runtime metadata such as:
- deployment mode
- deployment exposure
- auth readiness
- bootstrap state in authenticated deployments
- feature flags such as company deletion
curl http://localhost:3100/api/health
const res = await fetch("/api/health");
const health = await res.json();
import requests
health = requests.get("http://localhost:3100/api/health").json()
OpenAPI Specification
The server publishes a machine-readable description of its REST surface as an OpenAPI document:
GET /api/openapi.json
Point your own tooling — client generators, request validators, or an API explorer — at this endpoint to get the current set of routes, parameters, and schemas straight from the running instance. From the terminal, paperclipai access openapi fetches the same document for you. This reference is written for people; the OpenAPI document is the same surface in a form built for machines.
A Good First API Call
If you are already authenticated, a simple company read is a good way to verify access:
GET /api/companies/{companyId}
curl "http://localhost:3100/api/companies/company-1" \
-H "Authorization: Bearer <token>"
const res = await fetch("http://localhost:3100/api/companies/company-1", {
headers: {
Authorization: `Bearer ${token}`,
},
});
const company = await res.json();
import requests
company = requests.get(
"http://localhost:3100/api/companies/company-1",
headers={"Authorization": f"Bearer {token}"},
).json()
Reading This Reference
Use this page for the rules that apply everywhere. Then jump to the resource-specific page for the actual endpoint shapes, payloads, and examples:
If you are building against the API from code, the safest mental model is:
- pick the right caller identity first
- keep every request company-scoped where the route expects it
- treat
400,403,404,409, and422as meaningful business signals, not just transport errors