Integrations

Plug Claude, or any MCP client, into Choregos. Your agent lists your projects, opens and follows work items, and tells you what waits for your decision — with your rights, never more. It never decides: you do, in the console.

How agents, skills, connectors and clients fit together

Agent: does the work of a workflow step: instructions, a model, the skills it carries and the tools it may call. An internal agent runs in Choregos, under a budget; an external agent is an AI client registered here, whose calls carry its name and limits.
Skill: know-how an agent carries — a SKILL.md and its files. A skill grants no right.
Connector: an installed instance of a system Choregos reaches OUT to: a directory, a device fleet, an MCP server. Each operation is allowed, needs an approval, or is forbidden. A catalogue lists what the platform offers (MCP servers, connector types); installing one from it makes a connector. A project's delivery tools — tracker, repository, CI, deployment — are set in the project's settings.
Client: an AI assistant that reaches IN to Choregos through its MCP door — Claude Code, Cursor, ChatGPT — with its human's access token: their rights, never more, fewer once registered as an external agent. It never decides.

A client or a tracker opens a work item
→
Its workflow hands each step to an agent
→
The agent calls connectors under the policy
→
A person approves what the policy asks

A supplier's agent that you call over MCP is a connector (an MCP server), not an agent of your registry.

Connect REST API

The door serves every project you can read; tools take a project argument.

MCP URL …

Paste into a terminal — the REST API needs a full-access token, created in administration

curl -H "Authorization: Bearer <full-access token>" …/api/v1/me

Which client reaches what

Client	Calls from	Signs in with	Here and now
Claude Code	your machine	MCP token	ready
Claude Desktop	your machine	MCP token	ready
claude.ai	the vendor's cloud	OAuth	needs a public HTTPS address and OAuth
Cursor	your machine	MCP token	ready
VS Code	your machine	MCP token	ready
ChatGPT	the vendor's cloud	OAuth	needs a public HTTPS address and OAuth
Other MCP client	your machine	MCP token	ready
CLI	your machine	full-access token	ready
REST API	your machine	full-access token	ready

When it does not connect

401 — the token is missing, revoked or expired. 403 insufficient_scope — a full-access token: create an MCP one here.
The console's 404 page — the ingress does not route /mcp to the API (chart 0.14 or later).
Plain HTTP — Claude Desktop's bridge needs --allow-http; claude.ai and ChatGPT refuse it.
An internal host — claude.ai and ChatGPT call from their vendor's cloud (Anthropic: 160.79.104.0/21) and cannot reach a host behind your VPN; Claude Code can.
A tool is missing — a read-only token, a role without the right, or a project whose tracker is external.

Your AI clients

Your access tokens for the MCP door, and the AI client behind each once it has called. Registering one makes it an external agent of the registry. Every token is listed, and revoked, in Admin › My access tokens.

No AI client

Connect Claude Code or another client above: it appears here once it has called.

Another agent, over A2A

The workflows that opt in (a2a: in their definition) are skills of this organisation's agent. A caller follows each task until a person decides in the console; it never decides itself.

Agent card …

JSON-RPC endpoint …

It opens the A2A door only: the REST API and the MCP door refuse it. Revoke it in Admin › My access tokens.

Connect an AI client

Assistants that reach in through the MCP door — they never decide. Connecting one here registers it as an external agent: its calls carry the agent's name and limits, within your rights, and you get its configuration once. To connect one simply as yourself, use the steps above.