Skip to content

Connect your agent

Time: about 5 minutes · You need: ParleML running (Run ParleML) and an agent client on your machine. You'll end with: your agent answering questions about your ParleML workspaces.

ParleML hosts its MCP server at /mcp on the same address as the app, so there is nothing to install. Your client needs two things: that address and, for most clients, a token.

1. Copy the address

In ParleML, open Settings → Access. Under Your own agent, the MCP server line shows the address, for example http://localhost:8000/mcp. Choose Copy.

2. Create a token

On the same page, under API tokens:

  1. Enter a Token name you'll recognise, such as cursor-laptop.
  2. Pick a Scope: analyst to read, compute to also run bounded checks, builder to also build in the workspaces you choose (Tokens & scopes).
  3. Choose Mint token, and copy the token. It is shown only once.

The page fills the token into the configurations under Your own agent. Codex and Gemini CLI read it from the PARLEML_API_TOKEN environment variable instead. Only an administrator creates API tokens; with sign-in off (open mode), everyone is one.

claude.ai and ChatGPT need no token: they sign in through OAuth, and you approve the access on ParleML's consent page.

3. Configure your client

Under Your own agent, pick your client. The page shows the configuration with the address and the token filled in, and where it goes. The guide for each client:

Client Guide
Claude Code Claude Code: one claude mcp add command
Cursor, VS Code, Codex, Gemini CLI IDEs and CLIs
Claude Desktop, claude.ai Claude Desktop & claude.ai
ChatGPT ChatGPT

Keep a real token out of files a repository might commit. Most clients can read it from an environment variable, PARLEML_API_TOKEN by convention.

4. Ask something

Start your client and ask:

What workspaces do I have in ParleML?

Use names, not IDs: the agent looks them up. For more to try, see Questions to ask.

Give it the playbook

Ask the agent to call get_playbook first. It explains how to do ML work well in ParleML. For the full skill pack, export it from Settings → Skill packs (Skills & packaging).

Which clients can reach a local ParleML

Web clients and remote connectors call /mcp from their own servers, so they cannot reach a ParleML on localhost. Desktop clients, IDEs and CLIs call from your machine.

Client ParleML on localhost ParleML on public HTTPS
Claude Code, Cursor, VS Code, Codex, Gemini CLI Yes Yes
Claude Desktop (local server) Yes Yes
Claude Desktop (remote connector) No: needs a tunnel or proxy Yes
claude.ai No: needs a tunnel or proxy Yes
ChatGPT No: needs a tunnel or proxy Yes

Never demo a web client against localhost. For a temporary public address, use the reverse proxy or tunnel of your choice. It must pass the Authorization header through and must not buffer streamable-HTTP responses; /mcp needs nothing else.

Questions to ask

An analyst token reads and reasons; a builder token also builds in its workspaces. Releases, approvals and deletions stay with people, in ParleML.

  • Get oriented. What workspaces and experiments do I have, and what is the most recent experiment about? · Which dataset feeds this experiment, and how do its result tables join back to it?
  • Rules. What rules did this experiment learn, and which fire most among the rows it flags?
  • Archetypes. What characterises each archetype, and which one is the largest?
  • Drivers and what-ifs. What are the top drivers of this model, and how does income affect the predicted probability? · Where does the model underperform by segment? · If the threshold were 0.7, how many rows would we flag, by region?
  • Clusters and anomalies. Show the most anomalous entities: what makes them stand out?
  • Forecasting. What's the forecast for series X over the next 12 periods, with intervals?
  • Recommendations and ranking. Which recommended items lift the most over popularity? · How does NDCG vary by group?
  • SQL. Join the predictions back to the training data and compare predicted and actual rates by region.

Next