data-science-mcp¶
Model training, evaluation, and evolution tools for agentic ML workflows — an MCP Server + A2A Agent for the agent-utilities ecosystem, with compute delegated to the Rust epistemic-graph engine.
Official documentation
This site is the canonical reference for data-science-mcp, maintained alongside
every release.
Overview¶
data-science-mcp exposes typed, deterministic MCP tools for the full model
lifecycle — load datasets, fit and cross-validate estimators, rank models on a
Pareto frontier, generate interpretability suites, and fine-tune open-weight models
with SFT / DPO / GRPO. It provides:
MLEngine— a stateful façade over theepistemic-graphRust compute engine (fit,predict,evaluate,cross_validate,load_dataset,split_dataset). All ML compute runs in the engine over its MessagePack/UDS protocol; there is no scikit-learn compute path.- Action-routed MCP tools across five domains — model training, model evolution, interpretability, data management, and quantitative finance — each independently togglable to control LLM context.
- An in-house training substrate (Wave C): a deterministic SFT/DPO/GRPO corpus and reward engine plus torch/PEFT gradient trainers, CPU-smoke-tested on a toy model.
- A bundled Pydantic-AI A2A agent that wraps the tool surface for the Agent Control Protocol and the Agent Web UI.
Explore the documentation¶
- Installation — pip, source, extras, and the prebuilt Docker image.
- Deployment — run the MCP server and A2A agent, Docker Compose, Caddy + Technitium.
- Usage — the MCP tools, the
MLEnginePython API, and the CLI. - Overview — ecosystem role, enterprise readiness, and the concept registry.
- Model Training — the SFT/DPO/GRPO corpus, reward engine, and gradient trainers.
- Concepts — the
CONCEPT:DSCI-*registry.
Quick start¶
Run it as a network server:
See Installation and Deployment for the full matrix (PyPI extras, Docker image, all transports, the A2A agent, reverse proxy, DNS).