The Metis MCP server

metis mcp serves the Model Context Protocol over stdio, so AI tools (Claude Code, Claude Desktop, anything MCP-capable) can build and run workflows against a RUNNING Metis instance. It is a thin client of the HTTP control plane - point it at metis up, docker compose, or a remote deployment.

Setup

{
  "mcpServers": {
    "metis": {
      "command": "npx",
      "args": ["@mindlynx/metis-cli", "mcp"],
      "env": { "METIS_URL": "http://localhost:3000" }
    }
  }
}

Auth: set METIS_TOKEN, or METIS_USER/METIS_SECRET (defaults to the local dev seed admin/metis). The server logs in lazily and re-logs-in on 401, and never writes anything except protocol frames to stdout.

Tools

Tool	What it does
list_node_types	Browse the node catalogue (searchable)
get_node_type	One type's config/output schema, long-form docs, and the branch handle ids its edges must use
list_connections	The saved connections (metadata only) - a connector node's config.connectorId points at one
list_workflows / get_workflow	Read what exists
validate_workflow	Dry-check a graph (start node, cycles, loop/branch rules) without saving
create_workflow / update_workflow	Save a definition (nodes + edges; the tool descriptions carry the exact shape)
publish_workflow	Make the current draft the live version a trigger can fire
delete_workflow	Soft-delete a workflow
run_workflow	Start a run and wait briefly; returns per-step outcomes
get_execution	A run's status + per-step outputs (loop children by their own execution id)
list_executions	Recent runs from Temporal visibility, filterable by status
manage_execution	Operate a run: cancel / terminate / reset
create_trigger	Bind a webhook / schedule / poll trigger so a published workflow fires on its own
list_triggers / delete_trigger	See and remove triggers
manage_schedule	Pause or resume a workflow's schedule

The catalogue tools return the same Guide documentation the inspector shows humans - so the model knows what a node is, how it works and how to wire its branches before it builds.

The full lifecycle

The tools cover a workflow end to end: browse node types and connections, validate a graph, create it, publish it, bind a trigger (webhook or cron), run it, and operate or reset the runs - all without leaving the AI tool. Building a connector workflow (SendGrid, Postgres) needs a connection first; list_connections shows the ids, but connections are created in the Metis UI so credentials never pass through the model.