Skip to content

ParleML agents: MCP, skills and sessions

In short

  • An agent (Claude, ChatGPT, Cursor, Codex, Gemini CLI and others) works in ParleML through its MCP server at /mcp. MCP (Model Context Protocol) is the standard way agents use tools.
  • You can bring your own agent: connect it in a few minutes and ask about your workspaces. Start with Connect your agent.
  • Or let ParleML run one for you in an agent session, inside a workspace, with decisions that come back to you. See Agent sessions.
  • What an agent may do is set by its token's scope: analyst reads, compute adds bounded calculations, builder builds in chosen workspaces. Promoting a model, deleting and changing settings always stay with people.

Two ways to work with an agent

Your own agent. Claude Desktop, claude.ai, Claude Code, ChatGPT, Cursor, VS Code, Codex or Gemini CLI connects to ParleML's MCP server and answers questions about your workspaces. It runs read-only SQL over the workspace data, reads model evaluations, insights, per-row explanations and charts, follows runs, compares experiments and runs bounded checks. With a builder token it also delivers models through ParleML's own engine: recipes, experiment plans, launches and model registration. The skill packs teach it the workflow.

Agent sessions. ParleML runs an agent CLI on its own server, signed in with your account, to do ML work in one workspace. It uses the same tools and skills, plus a second MCP server, the session MCP. Through it the session asks you structured questions (decisions), starts its steps and waits on jobs. With Decide automatically on, the session answers the decisions it is allowed to answer itself (the session MCP).

Details: why there are two MCP servers

The two servers stay apart. The business MCP (the one this site documents) holds every ML capability and no session logic. The session MCP holds the session's lifecycle (decisions, child steps, waits) and no ML. Your own agent never sees the session MCP: it asks you in chat instead.

ParleML computes, the agent reasons

Every number comes from ParleML's engines: ml4 (training: supervised, forecasting, unsupervised, association, signal check) and mli (evaluation, trust, fairness, rules, archetypes, EDA, model cards). Agents reason over those results. They never re-implement a number a tool returns.

Optional AI annotations add narrative, descriptions and suggestions, marked with who wrote them. They are not measurements, and every computed result is usable without them. The workspace profile governs them.

The three scopes

Scope The agent can
analyst (default) Read and reason: browse, read the catalog, run read-only SQL, profile, read EDA, evaluations, insights, charts and rules, follow runs and logs, compare experiments and models, read the platform's vocabulary (presets, templates, metrics), and read where a workspace stands (get_workspace_brief). Nothing changes.
compute The above, plus bounded single-shot calculations: probe_dataset (is there learnable signal, is the series forecastable, what does a trivial predictor score, how would a temporal split fall), what_if (score one hypothetical row), compare with kind="datasets" (drift against a reference) and simulate_decision (what each threshold, cost or capacity yields). Nothing changes.
builder The above, plus building in the workspaces the token is bound to, within a run budget: shape data (author_recipe, prepare_data), plan and launch experiments (run_experiment), score (score_batch) and register models (register_model).

Promoting a model, archiving or restoring a workspace, and deleting anything stay with people, in the ParleML app, for every token. Most tool results carry a ui_url link that opens the page showing the result. Tokens & scopes has the full rules.

Where to go

  1. Connect your agent: the few-minute path for each client.
  2. Agent sessions: start and follow a session, and configure agents.
  3. Tool catalog: every tool, its scope and what it covers.
  4. Playbook: the guidance the agent reads (get_playbook serves the same text).
  5. Security model: scopes, the builder's guard rails, the trust boundary.

Details: where ui_url links land

A ui_url opens the page that shows the result, on the tab that shows it when the result has its own tab. get_model_evaluation lands on the experiment's Performance tab, probe_dataset on the dataset's signal check, and get_run(experiment_id, view="log") on the run's console. Tools that read the object itself (get_resource, get_capabilities, what_if, get_run's summary) link its full /workspaces/<workspace>/… page, and compare links the workspace's compare page. The others know only the object's ID, so they link a short permalink (/datasets/<id>, /experiments/<id>, /models/<id> or /knowledge/<id>) that the app resolves to the same page.

Next