Skip to content

Claude Desktop & claude.ai

In short

  • Claude Desktop with a local ParleML: add a local (stdio) server to claude_desktop_config.json. It runs inside the ParleML container and works on localhost.
  • claude.ai, and Claude Desktop's remote connectors, call from Anthropic's servers. They need ParleML on public HTTPS, and sign in through OAuth, with no token to paste.
  • Claude Code has its own section: Claude Code.

Claude Desktop with a local ParleML

You need: ParleML running in Docker under the name parleml (Run ParleML), and a token from Settings → Access (Connect your agent).

  1. In Claude Desktop, open Settings → Developer → Edit Config. This opens claude_desktop_config.json.

  2. Add the parleml entry and save:

    {
  "mcpServers": {
    "parleml": {
      "command": "docker",
      "args": ["exec", "-i", "-e", "PARLEML_API_TOKEN=<your-token>", "parleml", "python", "-m", "parleml_mcp.server"]
    }
  }
}

    Replace <your-token> with your token. With sign-in off (open mode), you can leave out "-e", "PARLEML_API_TOKEN=<your-token>",. If the file already has other servers, add parleml next to them inside mcpServers.

  3. Quit Claude Desktop completely and start it again. The ParleML tools appear in its tools menu.

  4. Ask What workspaces do I have in ParleML?

Claude Desktop starts the MCP server inside the container, and the server calls ParleML from there, so localhost works. The config file holds the token: keep it private.

If the tools don't appear

Claude Desktop must find the docker command. If its log says it cannot start docker, put the full path in "command": on macOS usually /usr/local/bin/docker, on Windows C:\\Program Files\\Docker\\Docker\\resources\\bin\\docker.exe (backslashes doubled in JSON). Check that docker ps lists a container named parleml.

From a ParleML source checkout you can run the server with uv instead: see Local server (stdio).

claude.ai and Claude Desktop remote connectors

You need: ParleML reachable on public HTTPS (see which clients can reach a local ParleML).

  1. In claude.ai, open Settings → Connectors → Add custom connector. Claude Desktop has the same menu.
  2. Enter your ParleML MCP address, such as https://parleml.example.com/mcp.
  3. claude.ai opens ParleML's consent page. Sign in if asked.
  4. Choose what to allow: read; read and check (compute); or, if you are a ParleML administrator, build in the workspaces you pick, within a run budget. Approve.

The connection then shows under Settings → Access → Connected agents in ParleML. Choose Disconnect there to end it at any time. Anthropic's connector guide covers the claude.ai side.

Details: OAuth behind a proxy

claude.ai discovers ParleML's authorization server from /mcp, registers itself and opens the consent page. If ParleML sits behind a proxy that rewrites the host, set PARLEML_OAUTH_ISSUER to ParleML's public origin so discovery points at it.

Your first questions

Ask Claude to find your workspace and call get_workspace_brief, then to read get_playbook for the workflow.

ParleML also registers seven prompts: explore_workspace, audit_model, investigate_anomalies, triage_dataset, deliver_model, prepare_training_table and improve_experiment. The last three guide builder work and need a builder token. Clients differ in whether they show prompts; see Resources & prompts.

Next