Agents as MCP servers

Every agent is also an MCP server, at https://run.camelai.com/v1/agents/<agent id>/mcp (Streamable HTTP). Connect Claude, Cursor or any MCP client to it, and the model there can talk to your agent. The server has one tool, message, which sends the agent a message and returns its reply. The agent's own tools stay its own.

claude mcp add --transport http support https://run.camelai.com/v1/agents/client_.../mcp
# Then run /mcp in Claude Code to sign in. Or skip signing in with a token:
claude mcp add --transport http support https://run.camelai.com/v1/agents/client_.../mcp --header "Authorization: Bearer art_..."

The agent's id (client_...) is agent.id in the SDKs, and GET /v1/agents lists every agent with its key.

The message tool

The tool's description tells the calling model what the agent is, so it knows when to use it.

How it behaves

• One conversation. A message joins the agent's history exactly as a prompt (POST /v1/agents/{id}/prompt) does, next to messages from the API, the console and channels. There are no separate threads per caller; to give each caller its own conversation, give each one its own agent. Each message is a run, limited and billed like any prompt.
• Long turns. A call lasts as long as the turn. It sends progress notifications while the agent works (its text as it streams, each tool it uses, and a heartbeat every 20 seconds), so clients that reset their timeout on progress keep waiting. A caller that goes away doesn't stop the turn.
• Retries. Pass requestId to make a retry safe. A call your client sends again in the same session is recognized without one.
• A busy agent. A message to an agent that is working queues behind its current turn, as a prompt does. Once 32 requests are open, the call fails with The agent is busy: ....
• Errors. A failed run, a spend limit, a model error or a refused request comes back as a tool error carrying camelRun's message.

Human input

When the turn waits on a person and the client supports MCP elicitation, the question, approval, form or page comes up in the client, and the answer resumes the turn.

Otherwise the call ends with a result listing what the agent waits on (structuredContent.status is input_required) and how to answer it. Once it's answered, calling again with the same text and requestId returns the reply. A new message instead sets the waiting inputs aside.

Who may call it

Anyone who may prompt the agent: the agent's own token (the token it was created with), an API key of your account, or an OAuth sign-in. A 401 names the agent's protected-resource metadata, so clients sign in the same way as with the hosted MCP server. Other accounts' tokens get a 404.