managed-agent-control-mcp

Start, observe, and interact with Claude Managed Agents from any MCP client (Claude.ai, Claude Code, Cursor, mcp-remote, your own agent, …).

Claude Managed Agents run on Anthropic's platform — each is an agent definition (model, system prompt, tools, skills) that executes inside a sandbox environment. This MCP server puts a remote control on those agents: connect it to any MCP client and you can launch an agent, watch what it does, reply to it, approve the tools it wants to run, and stop it. Claude.ai is the showcase client (do it all from a normal conversation), but the same tools work from Claude Code, Cursor, mcp-remote, the MCP Inspector, or a custom MCP client. It is the runtime companion to the Terraform provider that defines agents declaratively.

It runs as a local stdio server, a generic HTTP container, or on AWS Lambda, with pluggable inbound auth (static bearer, generic OIDC, or an AWS Cognito preset).

[!NOTE] This is a community project. It is not maintained by, endorsed by, or affiliated with Anthropic. The Managed Agents API is a beta API (managed-agents-2026-04-01).

How it works

MCP client ──MCP──▶ managed-agent-control-mcp ──HTTPS (x-api-key)──▶ Managed Agents API
(Claude.ai,         (this server: tools + auth)                     (agents run here)
 Claude Code, …)

You drive a loop: discover an agent → start a session → observe by polling events → interact (reply / interrupt / approve tools) → end. Because MCP tool calls are request/response, observation is by polling (session_events / session_get) — there is no live stream into the chat.

Quickstart (local, <2 min)

You need an ANTHROPIC_API_KEY with Managed Agents access and uv.

git clone https://github.com/modus-agendi/managed-agent-control-mcp
cd managed-agent-control-mcp
uv sync

# Run the server over stdio:
ANTHROPIC_API_KEY=sk-ant-... uv run python -m managed_agents_mcp

Explore the tools interactively with the MCP Inspector:

ANTHROPIC_API_KEY=sk-ant-... npx @modelcontextprotocol/inspector \
  uv run python -m managed_agents_mcp

Register it with Claude Code (a project-scoped .mcp.json is included):

claude mcp add managed-agent-control -- uv run python -m managed_agents_mcp

Once published to PyPI, you can skip the clone: uvx managed-agent-control-mcp.

Tools

See docs/tools.md for every argument and return shape.

Configuration

Only ANTHROPIC_API_KEY is required for local use. Full reference in docs/configuration.md.

Authentication

There are two auth layers (kept separate by design):

• Outbound — how this server calls Anthropic: your ANTHROPIC_API_KEY.
• Inbound — how MCP clients authenticate to this server. Pluggable and required for any HTTP deployment:
  • bearer — a shared static token. Simplest; works with Claude.ai connectors and mcp-remote.
  • oidc — verify JWTs from any OIDC provider (Auth0, Okta, Keycloak, Entra, Cognito).
  • cognito — the OIDC verifier plus the hosted-UI OAuth facade Cognito needs.

Setup and Claude.ai connector onboarding: docs/authentication.md.

[!WARNING] Never expose an HTTP deployment without an inbound auth mode set. Anyone who passes inbound auth can drive your Anthropic key — scope the key and use the guardrails.

Deployment

Details: docs/deployment.md.

Documentation

• Configuration — every env var + guardrails
• Authentication — bearer / OIDC / Cognito + Claude.ai setup
• Deployment — local, container, Lambda
• Architecture — module map, the two auth layers, the polling model
• Tools — full tool reference
• Examples — end-to-end deployment walkthroughs (host + OAuth provider)

Contributing

Contributions welcome — see CONTRIBUTING.md and the Code of Conduct. Found a security issue? See SECURITY.md.

License

MIT — see LICENSE. Maintained by Andrei Svirida.