predictlm-mcp 0.1.0

MCP server exposing PredictLM tabular foundation models as a tool for LLM agents (Claude Desktop, Cursor, Continue, etc.)

predictlm-mcp

MCP server exposing the PredictLM tabular foundation model as a tool for LLM agents.

Give Claude Desktop, Cursor, Continue.dev, Cody — any MCP client — a calibrated tabular-prediction tool. Drop a CSV in chat, ask "predict who'll churn", get answers back with uncertainty bars. No fine-tuning, no per-dataset training, no hallucinated numbers.

Status: alpha (0.1.0). API likely to evolve based on early-adopter feedback.

Install

pip install predictlm-mcp

Use with Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "predictlm": {
      "command": "predictlm-mcp",
      "args": ["--model", "predictlm-mini-13m", "--device", "cpu"]
    }
  }
}

Restart Claude Desktop. The tabular_predict and tabular_describe tools are now available to Claude.

Use with Cursor / Continue.dev

Both clients support MCP servers via their MCP config files — point the command at predictlm-mcp.

Tools

tabular_predict

Predict labels for query rows given a context of labeled rows.

Inputs:

• context_table (CSV string OR array of objects) — labeled rows including the target column
• target_column (string) — name of the column to predict
• query_table (CSV string OR array of objects) — unlabeled rows to predict
• task_type (auto | classification | regression, default auto)
• return_uncertainty (bool, default true)

Returns:

{
  "task_type": "classification",
  "predictions": ["churn", "stay", "churn", ...],
  "class_probabilities": [[0.82, 0.18], ...],
  "uncertainty": [0.18, 0.45, 0.12, ...],
  "context_n_rows": 200,
  "query_n_rows": 50,
  "model": {"name": "predictlm-mini-13m", "params": 13500000, "license": "Apache-2.0"}
}

tabular_describe

Inspect the loaded model: name, params, license, supported tasks, recommended context size.

Example agent transcript

User: I dropped customers.csv. Predict next-month churn for the 50 rows in
      new_customers.csv. Use the first 200 rows of customers.csv as context.

Claude:  (calls tabular_describe — confirms PredictLM-Mini, 13M params)
         (calls tabular_predict with context_table=customers.csv[:200],
          target_column="churn", query_table=new_customers.csv)

Claude:  Of 50 new customers, PredictLM predicts 12 are likely to churn next
         month with high confidence (≥0.8 probability), 21 are uncertain
         (0.4–0.8), and 17 are likely to stay. The 12 high-confidence churn
         predictions share a tenure < 4 months and >2 support tickets in the
         last 30 days — that's the dominant signal in the context data.

This is the difference between an LLM hallucinating predictions and an LLM delegating prediction to a calibrated model and reasoning about the result.

Why PredictLM specifically?

• Apache-2.0 open weights — commercial-OK, on-device, HIPAA/SOC2 compliant by deployment
• Calibrated uncertainty — BarDistribution head returns predictive distributions, not just point predictions
• 13M / 26M parameters — runs on a laptop (no GPU required)
• In-context learning — no per-dataset training; just label some context rows
• Distilled from a 26M Base — proven recipe, eval'd on 25 OpenML datasets

License

Apache-2.0. Free for commercial use, including in agent products.

Links

• Homepage: https://zerooneresearch.ai/predictlm/
• Model card (Mini): https://huggingface.co/zerooneresearch/predictlm-mini-13m
• Model card (Base): https://huggingface.co/zerooneresearch/predictlm-base-26m
• Issues: https://github.com/matej-01RAI/predictlm-mcp/issues