Data Science Mcp

CLI or API | MCP | Agent

Version: 2.2.0

Documentation — Installation, deployment, usage across the MCP, Python API, and CLI interfaces, and the in-house model-training substrate are maintained in the official documentation.

Overview

Data Science Mcp is a production-grade Agent and Model Context Protocol (MCP) server designed to interface directly with Data Science MCP Server — Model training, evaluation, and evolution tools for agentic ML workflows. Integrates with agent-connector-sdk IModelEvolver (CONCEPT:AU-AHE.harness.self-improvement-overview)..

Key Features

• Consolidated Action-Routed MCP Tools: Minimizes token overhead and eliminates tool bloat in LLM contexts by grouping methods into optimized, togglable tool modules.
• Enterprise-Grade Security: Complete support for Eunomia policies, OIDC token delegation, and granular execution context tracking.
• Integrated Graph Agent: Built-in Pydantic AI agent supporting the Agent Control Protocol (ACP) and standard Web interfaces (AG-UI).
• Native Telemetry & Tracing: Out-of-the-box OpenTelemetry exports and native Langfuse tracing.
• In-House Model Training (Wave C): A deterministic SFT/DPO/GRPO corpus + reward engine plus torch/PEFT gradient trainers (LoRA/QLoRA, TIES adapter merge, vLLM rollouts, checkpoint→reliability-suite eval hooks). Loss/optimizer kernels are CPU-smoke-tested on a toy model; real fine-tunes run on the GB10. Install with pip install data-science-mcp[training]. MCP tools: build_training_dataset, compose_reward, train_sft, train_dpo, train_grpo, merge_adapters_ties. See docs/training.md.

CLI or API

This agent wraps the Data Science MCP Server — Model training, evaluation, and evolution tools for agentic ML workflows. Integrates with agent-connector-sdk IModelEvolver (CONCEPT:AU-AHE.harness.self-improvement-overview). API. The operator can interact with it programmatically or via its integrated execution entrypoints.

Detailed instructions on how to use the underlying API wrappers, extended schema bindings, and developer SDK references are maintained in docs/index.md.

MCP

This server use dynamic Action-Routed tools to optimize token overhead and maximize IDE compatibility.

Available MCP Tools

Auto-generated — do not edit (synced by the mcp-readme-table pre-commit hook).

Condensed action-routed tools (MCP_TOOL_MODE=condensed)

Verbose 1:1 API-mapped tools (MCP_TOOL_MODE=verbose or both)

38 action-routed tool(s) · 9 verbose 1:1 tool(s). Each is enabled unless its <DOMAIN>TOOL toggle is set false; MCP_TOOL_MODE selects the surface (intent default — the six verb-tools, granular set loaded on demand · condensed action-routed · verbose 1:1 · both). Auto-generated — do not edit.

Detailed tool schemas, parameter shapes, and validation constraints are preserved in docs/usage.md.

Dynamic Tool Selection & Visibility

This MCP server supports dynamic toolset selection and visibility filtering at runtime. This allows the operator to restrict the set of exposed tools to prevent blowing up the LLM's context window.

The operator can configure tool filtering via multiple input channels:

• CLI Arguments: Pass --tools or --toolsets (or their disabled counterparts --disabled-tools and --disabled-toolsets) during startup.
• Environment Variables: Define standard environment variables:
  • MCP_ENABLED_TOOLS / MCP_DISABLED_TOOLS
  • MCP_ENABLED_TAGS / MCP_DISABLED_TAGS
• HTTP SSE Request Headers: Pass custom headers during transport initialization:
  • x-mcp-enabled-tools / x-mcp-disabled-tools
  • x-mcp-enabled-tags / x-mcp-disabled-tags
• HTTP SSE Request Query Parameters: Append query parameters directly to the operator's transport connection URL:
  • ?tools=tool1,tool2
  • ?tags=tag1

When query strings or parameters are supplied, an LLM-free Knowledge Graph resolution layer (using DynamicToolOrchestrator) matches query intents against known tool tags, names, or descriptions, with safe fallback and automated 24-hour background cache refreshing.

MCP Configuration Examples

Install the connector-focused [mcp] extra. Examples use data-science-mcp[mcp] to add FastMCP / FastAPI through agent-connector-sdk[mcp]; the required Agent Utilities core still carries epistemic-graph[full]. The [agent-runtime] extra additionally enables model orchestration.

stdio Transport (local IDEs — Cursor, Claude Desktop, VS Code)

{
  "mcpServers": {
    "data-science-mcp": {
      "command": "uvx",
      "args": [
        "--from",
        "data-science-mcp[mcp]",
        "data-science-mcp"
      ],
      "env": {
        "MCP_TOOL_MODE": "intent",
        "DATA_ENGINETOOL": "True",
        "DATA_MANAGEMENTTOOL": "True",
        "DATA_SCIENCE_KERNEL_CONTAINER_RUNTIME": "docker",
        "DATA_SCIENCE_KERNEL_SANDBOX_IMAGE": "registry.example/image@sha256:<digest>",
        "DEEP_DELEGATETOOL": "True",
        "DSM_NEAR_PAIRS_LOCAL_MAX": "20000",
        "INFERENCE_API_KEY": "EMPTY",
        "INFERENCE_BACKEND": "vllm",
        "INTERPRETABILITYTOOL": "True",
        "KERNEL_SPECIALIZETOOL": "True",
        "MODEL_EVOLUTIONTOOL": "True",
        "MODEL_TRAININGTOOL": "True",
        "QUANTTOOL": "True",
        "TRAINERTOOL": "True",
        "TRAINING_DATATOOL": "True"
      }
    }
  }
}

Runtime references require an alias-aware launcher such as Arbora. Other launchers must omit those entries and inject the resolved values through their own runtime secret boundary.

Streamable-HTTP Transport (networked / production)

{
  "mcpServers": {
    "data-science-mcp": {
      "command": "uvx",
      "args": [
        "--from",
        "data-science-mcp[mcp]",
        "data-science-mcp",
        "--transport",
        "streamable-http",
        "--port",
        "8000"
      ],
      "env": {
        "TRANSPORT": "streamable-http",
        "HOST": "127.0.0.1",
        "PORT": "8000",
        "MCP_TOOL_MODE": "intent",
        "DATA_ENGINETOOL": "True",
        "DATA_MANAGEMENTTOOL": "True",
        "DATA_SCIENCE_KERNEL_CONTAINER_RUNTIME": "docker",
        "DATA_SCIENCE_KERNEL_SANDBOX_IMAGE": "registry.example/image@sha256:<digest>",
        "DEEP_DELEGATETOOL": "True",
        "DSM_NEAR_PAIRS_LOCAL_MAX": "20000",
        "INFERENCE_API_KEY": "EMPTY",
        "INFERENCE_BACKEND": "vllm",
        "INTERPRETABILITYTOOL": "True",
        "KERNEL_SPECIALIZETOOL": "True",
        "MODEL_EVOLUTIONTOOL": "True",
        "MODEL_TRAININGTOOL": "True",
        "QUANTTOOL": "True",
        "TRAINERTOOL": "True",
        "TRAINING_DATATOOL": "True"
      }
    }
  }
}

Alternatively, connect to a pre-deployed Streamable-HTTP instance by url:

{
  "mcpServers": {
    "data-science-mcp": {
      "url": "http://localhost:8000/data-science-mcp/mcp"
    }
  }
}

Run a reviewed container image as a least-privilege stdio child (no listener or published port):

docker run -i --rm \
  --read-only \
  --cap-drop=ALL \
  --security-opt=no-new-privileges \
  --pids-limit=256 \
  --tmpfs /tmp:rw,noexec,nosuid,nodev,size=64m \
  -e TRANSPORT=stdio \
  -e MCP_TOOL_MODE=intent \
  -e DATA_ENGINETOOL=True \
  -e DATA_MANAGEMENTTOOL=True \
  -e DATA_SCIENCE_KERNEL_CONTAINER_RUNTIME=docker \
  -e DATA_SCIENCE_KERNEL_SANDBOX_IMAGE="registry.example/image@sha256:<digest>" \
  -e DEEP_DELEGATETOOL=True \
  -e DSM_NEAR_PAIRS_LOCAL_MAX=20000 \
  -e INFERENCE_API_KEY=EMPTY \
  -e INFERENCE_BACKEND=vllm \
  -e INTERPRETABILITYTOOL=True \
  -e KERNEL_SPECIALIZETOOL=True \
  -e MODEL_EVOLUTIONTOOL=True \
  -e MODEL_TRAININGTOOL=True \
  -e QUANTTOOL=True \
  -e TRAINERTOOL=True \
  -e TRAINING_DATATOOL=True \
  registry.example.invalid/data-science-mcp@sha256:<digest> data-science-mcp

For containerized network HTTP, supply an authenticated TLS ingress (or direct server TLS), exact MCP_ALLOWED_HOSTS, and an exact trusted-proxy CIDR policy through the operator-owned deployment profile. The generator does not emit an unauthenticated non-loopback listener.

Auto-generated from the code-read env surface (MCP_TOOL_MODE + package vars) — do not edit.

Additional Deployment Options

data-science-mcp can run as a local stdio process or container, or behind a remote network boundary. The Deployment guide carries the detailed transport contract.

• Local container — launch a reviewed immutable image as a least-privilege stdio child with no listener or published port.
• Remote URL — connect through an operator-supplied authenticated HTTPS ingress. Keep its URL, outbound identity references, trust profile, and exact MCP_ALLOWED_HOSTS in AgentConfig.

Environment Variables

Package environment variables

Inherited agent-connector-sdk variables (apply to every connector)

39 package + 16 inherited variable(s). Auto-generated from .env.example + the shared agent-connector-sdk set — do not edit.

Every variable the server reads, grouped by purpose.

MCP server / transport

Connection

Training / inference backend (full [training] extra)

Tool toggles

Each action-routed tool can be disabled individually via its toggle env var (set to false). The full list is in the Available MCP Tools table above.

Telemetry & governance

Agent CLI (full [agent] runtime only)

See .env.example for a copy-paste starting point.

Agent

This repository features a fully integrated Pydantic AI Graph Agent. It communicates over the Agent Control Protocol (ACP) and interacts smoothly with the Agent Web UI (AG-UI) and Terminal interface.

Running the Agent CLI

To start the interactive command-line agent:

# Set credentials
export DATA_SCIENCE_MCP_URL="your_value"
export DATA_SCIENCE_MCP_TOKEN="your_value"

# Run the agent server
data-science-agent --provider openai --model-id gpt-4o

Docker Compose Orchestration

The following docker/agent.compose.yml configures the Agent, Web UI, and Terminal Interface together:

version: '3.8'

services:
  data-science-mcp-mcp:
    image: example/data-science-mcp:mcp
    container_name: data-science-mcp-mcp
    hostname: data-science-mcp-mcp
    restart: always
    env_file:
      - ../.env
    environment:
      - PYTHONUNBUFFERED=1
      - HOST=0.0.0.0
      - PORT=8000
      - TRANSPORT=streamable-http
    ports:
      - "8000:8000"
    healthcheck:
      test: ["CMD", "python3", "-c", "import urllib.request; urllib.request.urlopen('http://localhost:8000/health')"]
      interval: 30s
      timeout: 10s
      retries: 3
      start_period: 10s
    logging:
      driver: json-file
      options:
        max-size: "10m"
        max-file: "3"

  data-science-mcp-agent:
    image: example/data-science-mcp@sha256:<digest>
    container_name: data-science-mcp-agent
    hostname: data-science-mcp-agent
    restart: always
    depends_on:
      - data-science-mcp-mcp
    env_file:
      - ../.env
    command: [ "data-science-agent" ]
    environment:
      - PYTHONUNBUFFERED=1
      - HOST=0.0.0.0
      - PORT=9004
      - MCP_URL=http://data-science-mcp-mcp:8000/mcp
      - PROVIDER=${PROVIDER:-openai}
      - MODEL_ID=${MODEL_ID:-gpt-4o}
      - ENABLE_WEB_UI=True
      - ENABLE_OTEL=True
    ports:
      - "9004:9004"
    healthcheck:
      test: ["CMD", "python3", "-c", "import urllib.request; urllib.request.urlopen('http://localhost:9004/health')"]
      interval: 30s
      timeout: 10s
      retries: 3
      start_period: 10s
    logging:
      driver: json-file
      options:
        max-size: "10m"
        max-file: "3"

Detailed graph node architecture explanations, custom skill configurations, and agentic trace guides are available in docs/deployment.md.

Security & Governance

Built directly upon the enterprise-ready agent-connector-sdk core, standard security parameters are fully supported:

Access Control & Policy Enforcement

• Eunomia Policies: Fine-grained, policy-driven tool authorization. Supports none, local embedded (mcp_policies.json), or centralized remote modes.
• OIDC Token Delegation: Compliant with RFC 8693 token exchange for flowing authenticating user credentials from Web UI / ACP → Agent → MCP.
• Scoped Credentials: Execution context runs restricted to the specific caller identity.

Runtime Security Grid

Installation

Pick the extra that matches what the operator want to run:

Heavy ML extras are opt-in and imported lazily — add them only when needed: [training] (torch/PEFT gradient trainers), [training-scale] (DeepSpeed/FlashAttention, GPU-host only), [training-fast] (Liger Triton kernels), [datasets] (scikit-learn sample loaders), [eval] (LightEval), [tracking] (MLflow). See docs/training.md.

# Connector-focused MCP server (includes the shared graph engine)
uv pip install "data-science-mcp[mcp]"

# Agent runtime (adds model orchestration to the shared graph engine)
uv pip install "data-science-mcp[agent]"

# Everything (development)
uv pip install "data-science-mcp[all]"      # or: python -m pip install "data-science-mcp[all]"

Container images (:mcp vs :agent)

One multi-stage docker/Dockerfile builds two right-sized images, selected by --target:

docker build --target mcp   -t example/data-science-mcp:mcp    docker/   # connector-focused MCP server
docker build --target agent -t example/data-science-mcp:agent-local docker/   # agent runtime

docker/mcp.compose.yml runs the connector-focused :mcp server; docker/agent.compose.yml runs the agent (immutable agent digest) with a co-located :mcp sidecar.

Knowledge-graph database (epistemic-graph)

data-science-mcp uses the mandatory epistemic-graph[full] compute capabilities on top of the epistemic-graph[full] runtime carried by every Agent Utilities install. The connector-focused [mcp] surface uses that engine without enabling model orchestration; [agent] adds the orchestration stack. Local deployments can use the bundled engine. For production or shared state, configure a dedicated epistemic-graph service. See the epistemic-graph deployment guide.

Documentation

The complete documentation is published as the official documentation site and is the recommended reference for installation, deployment, and day-to-day operation.

Repository Owners

Contribute

Contributions are welcome! Please ensure code quality by running local checks before submitting pull requests:

• Format code using ruff format .
• Lint code using ruff check .
• Validate type-safety with mypy .
• Run test suites using pytest

Deploy with agent-connector-sdk-deployment

Provision this package with the consolidated agent-connector-sdk-deployment workflow. It selects an installed-package, editable-source, or immutable-container path; records only runtime secret and TLS-profile references in AgentConfig; and runs doctor, registration, policy, observability, and rollback gates. Ask your agent to "deploy data-science-mcp with agent-connector-sdk-deployment".

The repository embeds no deployment profile, credential value, certificate path, or environment-specific endpoint. Supply those at runtime through AgentConfig and the configured secret provider.

Governed capability contract

This package ships a compact canonical skill surface with specialist procedures kept as referenced workflows. The current MCP tools, skill metadata, connector_manifest.yml, ontology, mappings, shapes, fixtures, migrations, tool-schema fingerprints, and certification metadata form one versioned capability contract. Validate them together; do not rely on stale tool names or historical per-task skill wrappers.

Runtime endpoints, credentials, certificate trust, tenant identity, retention, and observability policy are deployment inputs and are never packaged values. See Configuration, trust, and privacy before enabling a network transport, connector ingestion, Arbora delegation, or trace export.