Deployment¶

Deployment Options¶

data-science-mcp supports local stdio, a loopback-only development listener, a least-privilege stdio container, and a remote authenticated HTTPS boundary. Provider endpoint, credential, selector, identity, and trust material are supplied at runtime through AgentConfig; none is stored in this repository.

Installed stdio process¶

{
  "mcpServers": {
    "data-science": {
      "command": "data-science-mcp",
      "args": [],
      "env": {"MCP_TOOL_MODE": "intent"}
    }
  }
}

Loopback development listener¶

data-science-mcp --transport streamable-http --host 127.0.0.1 --port 8000

Do not expose this listener beyond loopback. Network deployments require direct TLS or an explicitly trusted TLS-terminating ingress, configured authentication, exact MCP_ALLOWED_HOSTS, and an exact trusted-proxy CIDR policy.

Least-privilege local container¶

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 \
  registry.example.invalid/data-science-mcp@sha256:<digest> data-science-mcp

The operator projects the selected AgentConfig profile into the process at runtime; the image remains immutable and contains no environment connection profile.

Remote authenticated HTTPS endpoint¶

{
  "mcpServers": {
    "data-science": {"url": "https://service.example.invalid/mcp"}
  }
}

Store the real remote URL, outbound identity reference, and TLS-profile reference in AgentConfig, not in MCP client JSON or documentation.

This page covers running data-science-mcp as a long-lived server: the transports, a Docker Compose stack, the bundled A2A agent, putting it behind a Caddy reverse proxy, and giving it a DNS name with Technitium.

data-science-mcp ships both an MCP server (console script data-science-mcp) and an A2A agent server (console script data-science-agent). The MCP server is a typed, deterministic tool surface; the agent wraps it for the Agent Control Protocol and the Agent Web UI.

Run the MCP server¶

The transport is selected with --transport (or the TRANSPORT env var):

stdio (default)streamable-httpsse

data-science-mcp

For IDE / desktop MCP clients that launch the server as a subprocess.

data-science-mcp --transport streamable-http --host 0.0.0.0 --port 8000

A network server with a /health endpoint and /mcp route.

data-science-mcp --transport sse --host 0.0.0.0 --port 8000

Health check (HTTP transports):

curl -s http://localhost:8000/health        # {"status":"OK"}

Configuration (environment)¶

data-science-mcp is configured entirely from the environment. The required runtime settings:

Var	Default	Meaning
`HOST`	`0.0.0.0`	Bind address for HTTP transports
`PORT`	`8000`	Listen port for HTTP transports
`TRANSPORT`	`stdio`	`stdio`, `streamable-http`, or `sse`
`MODEL_TRAININGTOOL`	`True`	Register the model-training tool domain
`MODEL_EVOLUTIONTOOL`	`True`	Register the model-evolution tool domain
`INTERPRETABILITYTOOL`	`True`	Register the interpretability tool domain
`DATA_MANAGEMENTTOOL`	`True`	Register the data-management tool domain
`QUANTTOOL`	`True`	Register the quantitative-finance tool domain
`EPISTEMIC_GRAPH_SOCKET`	—	UDS path to the `epistemic-graph` compute engine
`EPISTEMIC_GRAPH_TCP`	—	TCP endpoint to the compute engine (alternative to the socket)

Telemetry (ENABLE_OTEL, OTEL_EXPORTER_OTLP_*) and access governance (EUNOMIA_TYPE, EUNOMIA_POLICY_FILE, EUNOMIA_REMOTE_URL) are optional. The full set, with defaults, is documented in .env.example. Copy it to .env and populate only what you use.

Docker Compose¶

The repo ships docker/mcp.compose.yml. It reads a sibling .env and publishes the HTTP server on :8000:

services:
  data-science-mcp-mcp:
    image: example/data-science-mcp@sha256:<digest>
    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

cp .env.example .env          # then edit as needed
docker compose -f docker/mcp.compose.yml up -d
docker compose -f docker/mcp.compose.yml logs -f

Run the A2A agent¶

data-science-mcp ships a bundled Pydantic-AI agent (console script data-science-agent) that connects to the MCP server and exposes the Agent Control Protocol plus the Agent Web UI. Install the agent extra and run it:

pip install "data-science-mcp[agent]"

data-science-agent \
  --provider openai --model-id gpt-4o \
  --host 0.0.0.0 --port 9004 \
  --mcp-url http://localhost:8000/mcp

The agent reads MCP_URL (the MCP server's /mcp route) to discover the tool surface, and PROVIDER / MODEL_ID for the backing LLM. The repo ships docker/agent.compose.yml, which runs the MCP server and the agent together — the agent depends on the MCP service and reaches it by container name on :9004:

services:
  data-science-mcp-mcp:
    image: example/data-science-mcp@sha256:<digest>
    hostname: data-science-mcp-mcp
    environment:
      - HOST=0.0.0.0
      - PORT=8000
      - TRANSPORT=streamable-http
    ports: ["8000:8000"]

  data-science-mcp-agent:
    image: example/data-science-mcp@sha256:<digest>
    depends_on: [data-science-mcp-mcp]
    command: ["data-science-agent"]
    environment:
      - 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
    ports: ["9004:9004"]

docker compose -f docker/agent.compose.yml up -d

Behind a Caddy reverse proxy¶

Expose the HTTP server on a hostname with automatic TLS. Add to your Caddyfile:

# Internal (self-signed) — homelab .example.invalid zone
data-science-mcp.example.invalid {
    tls internal
    reverse_proxy data-science-mcp-mcp:8000
}

# Public — automatic Let's Encrypt
data-science-mcp.example.com {
    reverse_proxy data-science-mcp-mcp:8000
}

Reload Caddy:

docker compose -f services/caddy/compose.yml exec caddy caddy reload --config /etc/caddy/Caddyfile

DNS with Technitium¶

Point the hostname at the host running Caddy. Via the Technitium API:

curl -s "http://technitium.example.invalid:5380/api/zones/records/add" \
  --data-urlencode "token=$TECHNITIUM_DNS_TOKEN" \
  --data-urlencode "domain=data-science-mcp.example.invalid" \
  --data-urlencode "zone=arpa" \
  --data-urlencode "type=A" \
  --data-urlencode "ipAddress=192.0.2.10" \
  --data-urlencode "ttl=3600"

…or add an A record data-science-mcp.example.invalid → <caddy-host-ip> in the Technitium web console (http://technitium.example.invalid:5380). The ecosystem technitium-dns-mcp automates this as a tool.

Register with an MCP client¶

Add to your client's mcp_config.json:

{
  "mcpServers": {
    "data-science-mcp": {
      "command": "uvx",
      "args": ["--from", "data-science-mcp", "data-science-mcp"],
      "env": {
        "TRANSPORT": "stdio"
      }
    }
  }
}

For a remote HTTP server, point the client at http://data-science-mcp.example.invalid/mcp instead.