data-science-mcp — Concept Overview¶

Category: Intelligence | Ecosystem Role: MCP Server + A2A Agent Built on agent-utilities — the unified AGI Harness.

Description¶

Data Science MCP Server — Model training, evaluation, and evolution tools for agentic ML workflows. Integrates with agent-utilities IModelEvolver (CONCEPT:AU-AHE.harness.self-improvement-overview).

Enterprise Readiness¶

All agents in the ecosystem inherit enterprise-grade infrastructure from agent-utilities:

Feature	Status	Source
JWT/OIDC Authentication	✅ Built-in	`agent-utilities[auth]` — Authlib JWKS + API key middleware
OpenTelemetry Instrumentation	✅ Built-in	`agent-utilities[logfire]` — OTLP export, FastAPI auto-instrumentation
HashiCorp Vault Integration	✅ Built-in	`agent-utilities[vault]` — `secret://`, `env://`, `vault://` URI schemes
Audit Logging	✅ Built-in	Append-only compliance trail with 30+ action types (CONCEPT:AU-OS.governance.wasm-micro-agent-sandbox)
Token Usage Analytics	✅ Built-in	4-bucket tracking with budget alerting (CONCEPT:AU-OS.governance.wasm-micro-agent-sandbox)
Prompt Injection Defense	✅ Built-in	25+ pattern scanner + jailbreak taxonomy (CONCEPT:AU-OS.config.secrets-authentication)
Guardrail Engine	✅ Built-in	Input/output interception with block/redact/warn (CONCEPT:AU-OS.governance.reactive-multi-axis-budget)
Action Execution Pipeline	✅ Built-in	Token, cost, duration, and node transition limits Dry-run / commit / rollback phases (CONCEPT:AU-ORCH.adapter.kg-graph-materialization)
Resource Scheduling	✅ Built-in	Priority queuing + preemption limits (CONCEPT:AU-OS.state.cognitive-scheduler-preemption)
Session Concurrency	✅ Built-in	Enqueue/reject/interrupt/rollback (CONCEPT:AU-OS.governance.reactive-multi-axis-budget)

LLM Trainer¶

Beyond engine-backed classical ML, this project is the agent-driven LLM trainer for the ecosystem — it can create, pretrain from random init, and fine-tune models: robust SFT/DPO/GRPO with precision/accumulation/clipping/scheduling/checkpoint-resume and FSDP+DeepSpeed scale-out (CONCEPT:AU-AHE.trainer.high-caliber-llm-trainer/005), a corpus curation engine (CONCEPT:DS-AHE.trainer.data-engine), pretraining from scratch with a trained BPE tokenizer (CONCEPT:DS-AHE.trainer.concept-2), MLflow + KG tracking (CONCEPT:DS-AHE.trainer.concept-3), and an agent workflow that runs the whole loop (CONCEPT:AU-AHE.trainer.concept-2). See Installation for the capability→dependency matrix and Model Training for the recipes.

Concept Registry¶

This project implements or inherits the following ecosystem concepts (full CONCEPT:DSCI-* + CONCEPT:ML-* registry in Concepts):

Concept ID	Description	Source
AU-AHE.trainer.high-caliber-llm-trainer … AU-AHE.trainer.concept-2	LLM trainer — hardening, curation, pretrain, tracking, scale-out, eval, agent workflow	this project (cross-repo)
DS-ECO.mcp.model-training	Model Training Operations (in-house training substrate)	this project
AHE-3.1	Training Substrate (reward / distillation)	`agent-utilities` (bridge)
AU-AHE.harness.self-improvement-overview	Agent-Interpretable Model Evolver	`agent-utilities` (inherited)
AU-AHE.harness.width-diverse-best-k	LLM-Graded Interpretability Tests	`agent-utilities` (inherited)
ECO-4.1	MCP & Universal Skills	`agent-utilities` (inherited)
KG-2.17	Model Display Optimization	`agent-utilities` (inherited)

📖 Full Registry: See the agent-utilities concept index for the complete 5-Pillar concept registry.

Architecture¶

This project follows the standardized agent-package pattern:

data-science-mcp/
├── data_science_mcp/        # Source code
│   ├── __init__.py
│   ├── agent_server.py      # Entry point (create_graph_agent_server)
│   ├── api_client.py        # REST/GraphQL API wrapper
│   └── mcp_server.py        # FastMCP tool definitions
├── tests/                   # Test suite
├── docs/                    # Documentation
├── pyproject.toml           # Package metadata
├── mcp_config.json          # MCP server configuration
├── main_agent.json          # Agent identity & system prompt
└── Dockerfile               # Container deployment

MCP Configuration¶

stdio Mode¶

{
  "mcpServers": {
    "data-science-mcp": {
      "command": "uv",
      "args": ["run", "--with", "data-science-mcp", "data-mcp"],
      "env": {}
    }
  }
}

Streamable HTTP Mode¶

data-mcp --transport streamable-http --port 8001