OctOpus MCP Server — Data Science for Claude Code & Cursor
OctOpus ships a Model Context Protocol (MCP) server, so your coding agent can do real data science. Point Claude Code, Cursor, or any MCP-compatible client at OctOpus and it gains tools to profile a dataset, discover the right modelling direction, launch an autonomous research run, check status, and pull deployable artifacts — without leaving your editor.
What the MCP server exposes
The OctOpus MCP server registers a focused toolset: health, profile_csv, discover_direction, start_research_run, get_run_status, get_artifact_url, bootstrap_project, and bootstrap_and_export_handoff. Your agent calls them like any other tool — so 'profile this CSV and train the best churn model' becomes a single instruction that the agent fulfils end-to-end through OctOpus.
Why MCP beats copy-pasting code
AI coding assistants are great at writing code but they don't run a real research loop, manage holdout validation, or ship a model. By wiring OctOpus in over MCP, the agent delegates the hard ML — model-family rotation, leakage probes, Optuna tuning, holdout scoring, deployment — to a system built for it, and gets back validated artifacts instead of an untested notebook.
Set up in minutes
Start the server with `python3 mcp/octopus_mcp_server.py`, or register it via the `.mcp.json` already in the repo. It works with Claude Code, Cursor, and other MCP clients. From there your agent can profile data, kick off runs, and fetch the model.pkl and prediction-API URL programmatically.
Key capabilities
- MCP tools: profile_csv, discover_direction, start_research_run, get_run_status, get_artifact_url, bootstrap_project.
- Works with Claude Code, Cursor, and any MCP-compatible agent.
- Delegates the full ML research loop — your agent gets validated models, not raw code.
- Register via the .mcp.json in the repo, or run the server directly.
- Pull deployable artifacts (model.pkl, prediction API URL) programmatically.
Frequently asked questions
What is the OctOpus MCP server?
It's a Model Context Protocol server that exposes OctOpus's data-science tools — profile_csv, discover_direction, start_research_run, get_run_status, get_artifact_url, bootstrap_project — to any MCP-compatible agent, so Claude Code or Cursor can run a full autonomous ML workflow without leaving the editor.
How do I connect it to Claude Code or Cursor?
Run `python3 mcp/octopus_mcp_server.py`, or register the server via the .mcp.json in the repo. Once connected, your agent sees the OctOpus tools and can call them like any other MCP tool.
What does my agent get back?
Validated artifacts, not just code: a model trained and scored on a holdout, the train.py that produced it, run status, and the URLs to download the model.pkl or call the hosted prediction API.
Is this different from the OctOpus web app?
It's the same engine, a different entry point. The web app is for people who want a browser UI; the MCP server is for agents and developers who want to drive OctOpus programmatically from their IDE or another AI agent.