Agent Planners

Product

Agent Planners now has an API and MCP server — run it from Claude, Codex, or any agent

TL;DR

Agent Planners now has an outbound integration surface: a public REST API and an MCP server, both authenticated with per-organization API keys. Point Claude Desktop or Claude Code at the MCP server, or call the REST API from Codex, a cron job, or your own script, and it can create a task, poll its status, and list your connected accounts — running on the same orchestrator, the same credits, and the same approval-gated writes as a task started by hand in the Workspace.

What shipped

Two ways in, one underlying engine: a REST API at `/api/v1` (create a task, check its status and report, list connected accounts) and an MCP server at `/api/mcp` exposing the same three actions as tools — `create_task`, `get_task`, `list_accounts`. Both are authenticated the same way: a per-organization API key you mint from the dashboard, sent as a Bearer token.

Works with Claude, Codex, or anything that speaks HTTP

MCP clients plug in directly — add the server URL and your key as a Bearer header in Claude Desktop or Claude Code's MCP config, and its three tools show up alongside your other MCP servers.

Everything else — Codex, a CI job, a cron script, another agent framework's tool layer — can drive the exact same functionality over the plain REST API. It's just an HTTP endpoint with a Bearer token; there's no MCP-specific requirement to use it.

REST vs MCP — same actions, two shapes

REST API (/api/v1)	MCP server (/api/mcp)
Best for	Codex, scripts, CI, anything HTTP-capable	Claude Desktop, Claude Code, other MCP clients
Create a task	POST /tasks	create_task tool
Check status / report	GET /tasks/:id	get_task tool
List connected accounts	GET /accounts	list_accounts tool
Auth	Authorization: Bearer	Authorization header in the MCP client config

Same guardrails, now callable from outside the app

This is an outbound door into the same harness the Workspace uses — not a shortcut around it. A task created over the API or MCP runs the normal orchestrator → specialists → approval-gate pipeline, so nothing about how carefully it behaves changes just because it came from Codex instead of a browser tab.

• Keys are org-scoped and stored as a hash — a key can only ever act as its own organization, and a database leak can't be replayed against the API.
• External tasks spend your organization's credits exactly like a task started in the Workspace.
• Writes still pass the org's approval and guardrail stack unless the task explicitly sets `autoApprove` — and even then, guardrail caps still apply.
• Keys are scoped (`tasks:write`, `tasks:read`, `accounts:read`), rate-limited per key, and revocable instantly from the dashboard.

Set it up

1. 1Open API & MCP in your dashboard and click "New key".
2. 2Label it (e.g. "Claude Desktop — my laptop" or "Codex CI job") and pick scopes — start with all three unless you know you want less.
3. 3Copy the raw key immediately — it's shown once and stored only as a hash after that.
4. 4For an MCP client: add the server URL with the key as a Bearer header in its MCP config. For anything else: call the REST base with `Authorization: Bearer `.
5. 5Revoke the key from the same page any time to cut off access immediately.

Frequently asked questions

Does this bypass the approval gate?

No. Tasks created over the API or MCP server run the same orchestrator and approval-gate pipeline as a task started in the Workspace — writes still pause for approval unless the task sets autoApprove, and guardrail caps still apply either way.

Can Codex use this, or only Claude?

Both. Claude Desktop and Claude Code (and other MCP clients) connect directly to the MCP server. Codex, a script, a CI job, or any other agent that can send an HTTP request can drive the identical functionality over the plain REST API with the same Bearer key.

Does this cost extra?

No separate pricing — tasks created externally consume your organization's normal credits, the same as a task run from the Workspace.

How is this different from the existing MCP docs?

The opposite direction. Adding your own MCP server lets the agent call out to your tools. This feature lets an external agent — Claude, Codex, a script — call in to Agent Planners and run tasks on your accounts.

Can I limit what a key can do?

Yes — each key is scoped to tasks:write, tasks:read and/or accounts:read at creation, and can be revoked instantly, independent of your other keys.