Skip to content

Use ParleML from Claude Code

Time: about 5 minutes · You need: ParleML running (Run ParleML) and Claude Code on your machine. · You'll end with: Claude Code reading your ParleML workspaces and, if you allow it, training and explaining models in them.

ParleML has a built-in MCP server (Model Context Protocol, the standard way agents use tools). Claude Code connects to it over HTTP, with a token you create. Your agent then works through ParleML's own tools: it reads datasets and results, runs read-only SQL, and with the right token, builds experiments. Every object it creates shows up in the app.

1. Create a token

In ParleML, open Settings → Access, then API tokens:

  1. Token name: something you'll recognise, such as claude-code-laptop.

  2. Scope: choose how much the agent may do.

    Scope The agent can
    analyst Read everything: data, EDA, results, explanations, run SQL that only reads
    compute The above, plus bounded computations such as what-if scoring
    builder The above, plus build: recipes, column roles, experiments, explanations, model registration

    No token can promote a model, delete anything or change settings. Those stay with people, in the app.

  3. For builder, pick the Workspaces it may build in.

  4. Create it, then copy the token. It is shown only once.

2. Connect Claude Code

Still on Settings → Access, under Your own agent, choose Claude Code. The page shows the command with your server's address and the new token filled in. Run it in a terminal:

claude mcp add --transport http parleml http://localhost:8000/mcp \
  --header "Authorization: Bearer <your-token>"

Add --scope user to make ParleML available in every project, not only the current folder.

Check it: start claude and type /mcp. parleml is listed as connected, with its tools.

3. Ask something

Start with a question about what is already there:

What workspaces do I have in ParleML? For the churn one, summarise the best experiment:
how good it is on held-out data and what drives churn.

With a builder token you can ask for work:

In the Customer churn workspace, train a baseline that predicts churn, compare it with the
no-skill baseline, and explain the top three drivers. Tell me before registering anything.

Open the app alongside: the experiments Claude Code launches appear under the workspace's Experiments, with the same evaluation and explanation pages as any other.

Give it the playbook

ParleML's MCP server ships a playbook of how to do ML work well there. Ask Claude Code to call get_playbook first. For the full skill pack (framing, data, baseline, iteration, challenge, release), export it from Settings → Skill packs with Export with APM.

Good to know

  • Keep the token out of repositories. claude mcp add stores it in your Claude Code settings, not in your project folder. Revoke a token any time under Settings → Access.
  • The agent is you, scoped. It sees the workspaces and data your token allows, and its runs count against the token's budget (runs at once, runs per day, largest run).
  • Other clients. The same page has ready configurations for Claude Desktop, Cursor, Codex, Gemini CLI and more. The agents documentation covers each one.

Next