Skip to content

Docs · Connect

Connect other agent clients, both ways

Two directions. Your agent client (OpenClaw, Cursor, or any client that speaks MCP) can use Project Orchestrator as a set of tools. Or Project Orchestrator can hand work to another agent. The first works with documented settings; the second now opens a session, but without the Project Orchestrator tools, and has not been tried with a real OpenClaw.

Read from the public documentation of each side on 8 October 2026; the opening without tools was read from the backend code on 17 October 2026. Nothing here was run against a real OpenClaw yet.

Direction 1

Your client uses Project Orchestrator

Documented, not yet tried with a real OpenClaw

Project Orchestrator ships an MCP server. A client that can launch a command, or reach a Streamable HTTP address, can call its tools: plans, tasks, notes, decisions, code search.

  1. Step 1

    On the same machine: launch the server

    Declare the mcp_server program in OpenClaw. The server reads Neo4j and Meilisearch from its environment, or from the config.yaml of Project Orchestrator when a variable is missing. OpenClaw advises against writing secrets in its own configuration, so keep passwords in config.yaml.

    Terminal

    openclaw mcp set project-orchestrator '{"command":"mcp_server","env":{"NEO4J_URI":"bolt://localhost:7687","MEILISEARCH_URL":"http://localhost:7700"}}'
  2. Step 2

    Check that it answers

    Saving a definition proves nothing: OpenClaw does not contact the server when you save. This command opens a real connection.

    Terminal

    openclaw mcp doctor project-orchestrator --probe
  3. Step 3

    From another machine: turn on the remote address

    The /mcp address is off by default. When enabled it sits behind the sign-in of Project Orchestrator. Put it behind TLS (a reverse proxy or a tunnel) and never expose it without one.

    config.yaml

    remote_mcp:  enabled: true            # or REMOTE_MCP_ENABLED=true  bind_addr: "127.0.0.1"   # behind a TLS reverse proxy or tunnel
  4. Step 4

    Declare it in OpenClaw with OAuth

    OpenClaw uses SSE unless you write the transport, so name streamable-http. With auth set to oauth, the login command opens the sign-in. This sign-in flow was not run end to end between the two.

    Terminal

    openclaw mcp set project-orchestrator '{"url":"https://your-host/mcp","transport":"streamable-http","auth":"oauth"}'openclaw mcp login project-orchestrator
  5. Step 5

    Any other client

    Cursor and the other clients that launch an MCP command use the same program and the same variables. Only the file where the declaration goes changes: use the one your client documents.

    mcp.json

    {  "mcpServers": {    "project-orchestrator": { "command": "mcp_server" }  }}

Limits to know

  • OpenClaw hides MCP tools in its minimal tool profile; use the coding or messaging profile.
  • If OpenClaw cannot start the server, it skips it with a growing delay (30 seconds, doubling up to 10 minutes) before trying again.
  • The remote address needs TLS and the remote_mcp switch; the local program needs neither.

Direction 2

Project Orchestrator hands work to OpenClaw

Opens without the Project Orchestrator tools; not tried with a real OpenClaw

Project Orchestrator can run an agent that speaks the Agent Client Protocol (ACP). OpenClaw offers one: openclaw acp, a bridge to its Gateway. The declaration below is how it is wired.

  1. Step 1

    Declare the agent on the server

    A request never carries a command line: the command is declared once, on the server, under a name. Prefer a token file to a token on the command line, which any process listing shows.

    env

    CHAT_PROVIDER_ACP_COMMANDS='{"openclaw":["openclaw","acp","--url","wss://gateway-host:18789","--token-file","/path/to/gateway.token"]}'
  2. Step 2

    Add a provider instance that uses it

    The instance is of kind acp and names the declared agent as its preset.

    JSON

    { "kind": "acp", "preset": "openclaw" }

What the session lacks

OpenClaw refuses any MCP server attached to a session. Project Orchestrator no longer fails the opening: when the agent answers that it takes no per-session MCP server, it asks once more without it and opens the session without the Project Orchestrator tools, and lists project_orchestrator_tools among the degraded features of the session so that the conversation says so. The agent can work, but it cannot call plans, tasks or notes from that session. Two openings on the same agent at once no longer mix up which one kept its tools. This comes from reading both sides and from tests against stand-in agents, not from a real OpenClaw.

Limits to know

  • OpenClaw ACP does not offer file reads and writes or a terminal to the client, and loading a past session is only partly supported.
  • The model cannot be chosen from Project Orchestrator: OpenClaw decides.
  • No ACP agent has ever been run with Project Orchestrator for real: only a stand-in program that replays recorded exchanges.