#Run through MCP

Complete project setup and install a compatible Workflow first. The client must be able to execute IntLoom's Agent task protocol, including claiming tasks, invoking business Tools, and submitting results. Connecting an MCP endpoint alone does not complete tasks automatically.

#Start the project host

Keep the generated workflows declarations and useMcpAgent: true. This is the default. Project llms configuration is not required for this entry.

intloom start
intloom doctor --execution agent_ide
intloom flows

flows is a CLI query: any readiness it reports is for the CLI entry. Use doctor --execution agent_ide for the MCP execution path. Loading availability and entry readiness are different checks.

#Connect Codex

intloom config codex

The command prints the configuration generated for this project. Merge it into the intended project's .codex/config.toml; the command does not edit IDE configuration. The snippet points to /mcp/codex and uses an authentication helper instead of embedding the local token. Enable the connection using the client's project trust and permission controls.

A generic MCP client uses the authenticated loopback /mcp endpoint. See MCP reference for client implementers. Treat connection metadata and task ownership tokens as private.

#Run and continue

Ask the client to list Workflows, select the intended flowName, and call flow once with your intent. It should continue according to each returned Run snapshot:

The client uses call_agent_tool for declared business Tools and complete_agent_call for the final structured result. Human interaction uses interact, or answer_ask after showing the actual question if native forms are unavailable. It must not invent answers or call flow again to continue a wait.

#Verify the result

intloom runs
intloom artifacts

Read the returned Run and relevant Artifacts/Records. Terminal queries work regardless of which entry created the Run. completed describes process completion; evaluate the stored business results separately.

Keep the host running during client Agent tasks. Reconnecting a client with its original claim identity can continue a live task, but restarting the host interrupts it. See Recovery.

To use a project model from MCP instead, set useMcpAgent: false, configure models, and restart. No automatic service-model fallback occurs.