TabICL MCP Server

State-of-the-art tabular ML inside Claude or ChatGPT — share a CSV, ask a question, get predictions, explanations, and a report. Free and self-hostable.

TabICL (Inria, ICML 2025/2026) is a tabular foundation model: it classifies and regresses on tabular data without any hyperparameter tuning, via in-context learning. This MCP server makes it usable from natural language in any MCP-compatible assistant or agent.

You:    Here's my customer spreadsheet — who is likely to churn, and why?
Claude: [load_data]  → 2,340 rows, 12 columns, target "churned" looks binary
        [evaluate]   → 89% accuracy, 0.93 ROC-AUC on held-out data
        [explain]    → top drivers: contract_type, tenure, monthly_charges
        [create_report] → here's a shareable HTML report with the details…

Tools

Data limits: ~10k rows pasted inline, ~50k via URL or file. TabICLv2 handles 2–100 features natively.

Try it in 60 seconds

Connect the public server to claude.ai (Pro/Max/Team plans):

1. Go to Settings → Connectors → Add custom connector
2. Name: tabicl · URL: https://gblayer-tabicl-mcp.hf.space/mcp
3. Leave the OAuth fields empty and click Add — that's it.

Then paste this into a new chat:

Load this CSV and tell me: can you predict the churned column? How accurate is it, what drives churn the most, and can you make me a report? https://raw.githubusercontent.com/gblayer/tabicl-mcp/main/examples/customer_churn.csv

You'll get honest held-out metrics, the churn drivers ranked, and a shareable report — see a sample of what the report looks like (source).

Use it

Option A — Local (free, private, fastest)

Requires Python ≥ 3.10. With uv: nothing to install, clients run uvx tabicl-mcp.

Claude Desktop — ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "tabicl": { "command": "uvx", "args": ["tabicl-mcp"] }
  }
}

Claude Code: claude mcp add tabicl -- uvx tabicl-mcp

Cursor / VS Code (.vscode/mcp.json or Cursor Settings → MCP):

{
  "servers": {
    "tabicl": { "type": "stdio", "command": "uvx", "args": ["tabicl-mcp"] }
  }
}

(Or pip install tabicl-mcp and use "command": "tabicl-mcp".)

With a local server, load_data accepts file paths — no pasting needed, and no data ever leaves your machine.

Option B — Deploy your own private server (free, for claude.ai / ChatGPT)

Who this is for: when you use the public URL above, your CSV data is processed on a server operated by this project. That's fine for demos and non-sensitive data — but if you're working with confidential data (customer lists, medical records, company financials), deploy your own copy instead: identical functionality, but your data only ever touches infrastructure you control. A second reason: the public Space is a single shared free instance — if it's busy or asleep you wait, while your own deployment serves only you.

Deploying your own free endpoint on HuggingFace Spaces:

1. Create a Space → Docker SDK → CPU basic (free).

2. HF Spaces reads its deployment settings from a YAML header at the very top of the Space's README.md — add this before pushing (only needed for the copy that lives on HuggingFace, not for using the server):

   ---
   title: TabICL MCP
   emoji: 🤖
   colorFrom: blue
   colorTo: green
   sdk: docker
   app_port: 7860
   pinned: false
   ---
   

3. Push this repo to the Space (the Dockerfile pre-downloads model checkpoints at build).

4. Your MCP endpoint: https://YOUR-SPACE.hf.space/mcp

Then connect:

• claude.ai — Settings → Connectors → Add custom connector → paste the URL.
• ChatGPT — enable Developer Mode (Settings, paid plans) → Apps & Connectors → add the URL.

Free Spaces sleep after inactivity — the first request after idle takes a minute or two while the Space wakes. Everything after that is fast.

Option C — Docker anywhere

docker build -t tabicl-mcp .
docker run -p 7860:7860 tabicl-mcp
# MCP endpoint: http://localhost:7860/mcp   ·   healthcheck: /health

Development

python -m venv .venv && .venv/bin/pip install -e ".[dev]"
.venv/bin/pytest -m "not slow"   # fast tests
.venv/bin/pytest                 # includes real-model tests (downloads checkpoints)

Layout: tabicl_mcp/server.py (MCP tools) · data.py (ingestion + cache) · ml.py (evaluate/predict/importance) · report.py (HTML reports).

MIT licensed. TabICL itself is by soda-inria (BSD-3). PRs welcome.