QT/superset2
1
0
Fork 0
mirror of https://github.com/apache/superset.git synced 2026-04-07 10:31:50 +00:00
Code Issues Packages Projects Releases Wiki Activity
Files
296bd7e56b58d3c8c5f920e996f45d3ccd753ee8
superset2/docs/developer_docs/extensions/mcp-server.md

20 KiB
Raw Blame History

title, hide_title, sidebar_position, version
title hide_title sidebar_position version
MCP Server Deployment & Authentication true 9 1

MCP Server Deployment & Authentication

Superset includes a built-in Model Context Protocol (MCP) server that lets AI assistants -- Claude, ChatGPT, and other MCP-compatible clients -- interact with your Superset instance. Through MCP, clients can list dashboards, query datasets, execute SQL, create charts, and more.

This guide covers how to run, secure, and deploy the MCP server.

flowchart LR
    A["AI Client<br/>(Claude, ChatGPT, etc.)"] -- "MCP protocol<br/>(HTTP + JSON-RPC)" --> B["MCP Server<br/>(:5008/mcp)"]
    B -- "Superset context<br/>(app, db, RBAC)" --> C["Superset<br/>(:8088)"]
    C --> D[("Database<br/>(Postgres)")]

Quick Start

Get the MCP server running locally and connect an AI client in three steps.

1. Start the MCP server

The MCP server runs as a separate process alongside Superset:

superset mcp run --host 127.0.0.1 --port 5008
Flag Default Description
--host 127.0.0.1 Host to bind to
--port 5008 Port to bind to
--debug off Enable debug logging

The endpoint is available at http://<host>:<port>/mcp.

2. Set a development user

For local development, tell the MCP server which Superset user to impersonate (the user must already exist in your database):

# superset_config.py
MCP_DEV_USERNAME = "admin"

3. Connect an AI client

Point your MCP client at the server. For Claude Desktop, edit the config file:

  • macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
  • Windows: %APPDATA%\Claude\claude_desktop_config.json
  • Linux: ~/.config/Claude/claude_desktop_config.json
{
  "mcpServers": {
    "superset": {
      "url": "http://localhost:5008/mcp"
    }
  }
}

Restart Claude Desktop. The hammer icon in the chat bar confirms the connection.

See Connecting AI Clients for Claude Code, Claude Web, ChatGPT, and raw HTTP examples.

Prerequisites

  • Apache Superset 5.0+ running and accessible
  • Python 3.10+
  • The fastmcp package (pip install fastmcp)

Authentication

The MCP server supports multiple authentication methods depending on your deployment scenario.

flowchart TD
    R["Incoming MCP Request"] --> F{"MCP_AUTH_FACTORY<br/>set?"}
    F -- Yes --> CF["Custom Auth Provider"]
    F -- No --> AE{"MCP_AUTH_ENABLED?"}
    AE -- "True" --> JWT["JWT Validation"]
    AE -- "False" --> DU["Dev Mode<br/>(MCP_DEV_USERNAME)"]

    JWT --> ALG{"MCP_JWT_ALGORITHM"}
    ALG -- "RS256 + JWKS" --> JWKS["Fetch keys from<br/>MCP_JWKS_URI"]
    ALG -- "RS256 + static" --> PK["Use<br/>MCP_JWT_PUBLIC_KEY"]
    ALG -- "HS256" --> SEC["Use<br/>MCP_JWT_SECRET"]

    JWKS --> V["Validate token<br/>(exp, iss, aud, scopes)"]
    PK --> V
    SEC --> V
    V --> UR["Resolve Superset user<br/>from token claims"]
    UR --> OK["Authenticated request"]
    CF --> OK
    DU --> OK

Development Mode (No Auth)

Disable authentication and use a fixed user:

# superset_config.py
MCP_AUTH_ENABLED = False
MCP_DEV_USERNAME = "admin"

All operations run as the configured user.

:::warning Never use development mode in production. Always enable authentication for any internet-facing deployment. :::

JWT Authentication

For production, enable JWT-based authentication. The MCP server validates a Bearer token on every request.

Option A: RS256 with JWKS endpoint

The most common setup for OAuth 2.0 / OIDC providers that publish a JWKS (JSON Web Key Set) endpoint:

# superset_config.py
MCP_AUTH_ENABLED = True
MCP_JWT_ALGORITHM = "RS256"
MCP_JWKS_URI = "https://your-identity-provider.com/.well-known/jwks.json"
MCP_JWT_ISSUER = "https://your-identity-provider.com/"
MCP_JWT_AUDIENCE = "your-superset-instance"

Option B: RS256 with static public key

Use this when you have a fixed RSA key pair (e.g., self-signed tokens):

# superset_config.py
MCP_AUTH_ENABLED = True
MCP_JWT_ALGORITHM = "RS256"
MCP_JWT_PUBLIC_KEY = """-----BEGIN PUBLIC KEY-----
MIIBIjANBgkqhkiG9w0BAQEFAAOCAQ8AMIIBCgKCAQEA...
-----END PUBLIC KEY-----"""
MCP_JWT_ISSUER = "your-issuer"
MCP_JWT_AUDIENCE = "your-audience"

Option C: HS256 with shared secret

Use this when both the token issuer and the MCP server share a symmetric secret:

# superset_config.py
MCP_AUTH_ENABLED = True
MCP_JWT_ALGORITHM = "HS256"
MCP_JWT_SECRET = "your-shared-secret-key"
MCP_JWT_ISSUER = "your-issuer"
MCP_JWT_AUDIENCE = "your-audience"

:::warning Store MCP_JWT_SECRET securely. Never commit it to version control. Use environment variables:

import os
MCP_JWT_SECRET = os.environ.get("MCP_JWT_SECRET")

:::

JWT claims

The MCP server validates these standard claims:

Claim Config Key Description
exp -- Expiration time (always validated)
iss MCP_JWT_ISSUER Token issuer (optional but recommended)
aud MCP_JWT_AUDIENCE Token audience (optional but recommended)
sub -- Subject -- primary claim used to resolve the Superset user

User resolution

After validating the token, the MCP server resolves a Superset username from the claims. It checks these in order, using the first non-empty value:

  1. subject -- the standard sub claim (via the access token object)
  2. client_id -- for machine-to-machine tokens
  3. payload["sub"] -- fallback to raw payload
  4. payload["email"] -- email-based lookup
  5. payload["username"] -- explicit username claim

The resolved value must match a username in the Superset ab_user table.

Scoped access

Require specific scopes in the JWT to limit what MCP operations a token can perform:

# superset_config.py
MCP_REQUIRED_SCOPES = ["mcp:read", "mcp:write"]

Only tokens that include all required scopes are accepted.

Custom Auth Provider

For advanced scenarios (e.g., a proprietary auth system), provide a factory function. This takes precedence over all built-in JWT configuration:

# superset_config.py
def my_custom_auth_factory(app):
    """Return a FastMCP auth provider instance."""
    from fastmcp.server.auth.providers.jwt import JWTVerifier
    return JWTVerifier(
        jwks_uri="https://my-auth.example.com/.well-known/jwks.json",
        issuer="https://my-auth.example.com/",
        audience="superset-mcp",
    )

MCP_AUTH_FACTORY = my_custom_auth_factory

Connecting AI Clients

Claude Desktop

Local development (no auth):

{
  "mcpServers": {
    "superset": {
      "url": "http://localhost:5008/mcp"
    }
  }
}

With JWT authentication:

{
  "mcpServers": {
    "superset": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote@latest",
        "http://your-superset-host:5008/mcp",
        "--header",
        "Authorization: Bearer YOUR_TOKEN"
      ]
    }
  }
}

Claude Code (CLI)

Add to your project's .mcp.json:

{
  "mcpServers": {
    "superset": {
      "type": "url",
      "url": "http://localhost:5008/mcp"
    }
  }
}

With authentication:

{
  "mcpServers": {
    "superset": {
      "type": "url",
      "url": "http://localhost:5008/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_TOKEN"
      }
    }
  }
}

Claude Web (claude.ai)

  1. Open claude.ai
  2. Click the + button (or your profile icon)
  3. Select Connectors
  4. Click Manage Connectors > Add custom connector
  5. Enter a name and your MCP URL (e.g., https://your-superset-host/mcp)
  6. Click Add

:::info Custom connectors on Claude Web require a Pro, Max, Team, or Enterprise plan. :::

ChatGPT

  1. Click your profile icon > Settings > Apps and Connectors
  2. Enable Developer Mode in Advanced Settings
  3. In the chat composer, press + > Add sources > App > Connect more > Create app
  4. Enter a name and your MCP server URL
  5. Click I understand and continue

:::info ChatGPT MCP connectors require a Pro, Team, Enterprise, or Edu plan. :::

Direct HTTP requests

Call the MCP server directly with any HTTP client:

curl -X POST http://localhost:5008/mcp \
  -H 'Content-Type: application/json' \
  -H 'Authorization: Bearer YOUR_JWT_TOKEN' \
  -d '{"jsonrpc": "2.0", "method": "tools/list", "id": 1}'

Deployment

Single Process

The simplest setup: run the MCP server alongside Superset on the same host.

flowchart TD
    subgraph host["Host / VM"]
        direction TB
        S["Superset<br/>:8088"] --> DB[("Postgres")]
        M["MCP Server<br/>:5008"] --> DB
    end
    C["AI Client"] -- "HTTPS" --> P["Reverse Proxy<br/>(Nginx / Caddy)"]
    U["Browser"] -- "HTTPS" --> P
    P -- ":8088" --> S
    P -- ":5008/mcp" --> M

superset_config.py:

MCP_SERVICE_HOST = "0.0.0.0"
MCP_SERVICE_PORT = 5008
MCP_DEV_USERNAME = "admin"  # or enable JWT auth

# If behind a reverse proxy, set the public-facing URL so
# MCP-generated links (chart previews, SQL Lab URLs) resolve correctly:
MCP_SERVICE_URL = "https://superset.example.com"

Start both processes:

# Terminal 1 -- Superset web server
superset run -h 0.0.0.0 -p 8088

# Terminal 2 -- MCP server
superset mcp run --host 0.0.0.0 --port 5008

Nginx reverse proxy with TLS:

server {
    listen 443 ssl;
    server_name superset.example.com;

    ssl_certificate     /path/to/cert.pem;
    ssl_certificate_key /path/to/key.pem;

    # Superset web UI
    location / {
        proxy_pass http://127.0.0.1:8088;
        proxy_set_header Host $host;
        proxy_set_header X-Real-IP $remote_addr;
    }

    # MCP endpoint
    location /mcp {
        proxy_pass http://127.0.0.1:5008/mcp;
        proxy_set_header Host $host;
        proxy_set_header X-Real-IP $remote_addr;
        proxy_set_header Authorization $http_authorization;
    }
}

Docker Compose

Run Superset and the MCP server as separate containers sharing the same config:

# docker-compose.yml
services:
  superset:
    image: apache/superset:latest
    ports:
      - "8088:8088"
    volumes:
      - ./superset_config.py:/app/superset_config.py
    environment:
      - SUPERSET_CONFIG_PATH=/app/superset_config.py

  mcp:
    image: apache/superset:latest
    command: ["superset", "mcp", "run", "--host", "0.0.0.0", "--port", "5008"]
    ports:
      - "5008:5008"
    volumes:
      - ./superset_config.py:/app/superset_config.py
    environment:
      - SUPERSET_CONFIG_PATH=/app/superset_config.py
    depends_on:
      - superset

Both containers share the same superset_config.py, so authentication settings, database connections, and feature flags stay in sync.

Multi-Pod (Kubernetes)

For high-availability deployments, configure Redis so that replicas share session state:

flowchart TD
    LB["Load Balancer"] --> M1["MCP Pod 1"]
    LB --> M2["MCP Pod 2"]
    LB --> M3["MCP Pod 3"]
    M1 --> R[("Redis<br/>(session store)")]
    M2 --> R
    M3 --> R
    M1 --> DB[("Postgres")]
    M2 --> DB
    M3 --> DB

superset_config.py:

MCP_STORE_CONFIG = {
    "enabled": True,
    "CACHE_REDIS_URL": "redis://redis-host:6379/0",
    "event_store_max_events": 100,
    "event_store_ttl": 3600,
}

When CACHE_REDIS_URL is set, the MCP server uses a Redis-backed EventStore for session management, allowing replicas to share state. Without Redis, each pod manages its own in-memory sessions and stateful MCP interactions may fail when requests hit different replicas.

Configuration Reference

All MCP settings go in superset_config.py. Defaults are defined in superset/mcp_service/mcp_config.py.

Core

Setting Default Description
MCP_SERVICE_HOST "localhost" Host the MCP server binds to
MCP_SERVICE_PORT 5008 Port the MCP server binds to
MCP_SERVICE_URL None Public base URL for MCP-generated links (set this when behind a reverse proxy)
MCP_DEBUG False Enable debug logging
MCP_DEV_USERNAME -- Superset username for development mode (no auth)

Authentication

Setting Default Description
MCP_AUTH_ENABLED False Enable JWT authentication
MCP_JWT_ALGORITHM "RS256" JWT signing algorithm (RS256 or HS256)
MCP_JWKS_URI None JWKS endpoint URL (RS256)
MCP_JWT_PUBLIC_KEY None Static RSA public key string (RS256)
MCP_JWT_SECRET None Shared secret string (HS256)
MCP_JWT_ISSUER None Expected iss claim
MCP_JWT_AUDIENCE None Expected aud claim
MCP_REQUIRED_SCOPES [] Required JWT scopes
MCP_JWT_DEBUG_ERRORS False Log detailed JWT errors server-side (never exposed in HTTP responses per RFC 6750)
MCP_AUTH_FACTORY None Custom auth provider factory (flask_app) -> auth_provider. Takes precedence over built-in JWT

Response Size Guard

Limits response sizes to prevent exceeding LLM context windows:

MCP_RESPONSE_SIZE_CONFIG = {
    "enabled": True,
    "token_limit": 25000,
    "warn_threshold_pct": 80,
    "excluded_tools": [
        "health_check",
        "get_chart_preview",
        "generate_explore_link",
        "open_sql_lab_with_context",
    ],
}
Key Default Description
enabled True Enable response size checking
token_limit 25000 Maximum estimated token count per response
warn_threshold_pct 80 Warn when response exceeds this percentage of the limit
excluded_tools See above Tools exempt from size checking (e.g., tools that return URLs, not data)

Caching

Optional response caching for read-heavy workloads. Requires Redis when used with multiple replicas.

MCP_CACHE_CONFIG = {
    "enabled": False,
    "CACHE_KEY_PREFIX": None,
    "list_tools_ttl": 300,       # 5 min
    "list_resources_ttl": 300,
    "list_prompts_ttl": 300,
    "read_resource_ttl": 3600,   # 1 hour
    "get_prompt_ttl": 3600,
    "call_tool_ttl": 3600,
    "max_item_size": 1048576,    # 1 MB
    "excluded_tools": [
        "execute_sql",
        "generate_dashboard",
        "generate_chart",
        "update_chart",
    ],
}
Key Default Description
enabled False Enable response caching
CACHE_KEY_PREFIX None Optional prefix for cache keys (useful for shared Redis)
list_tools_ttl 300 Cache TTL in seconds for tools/list
list_resources_ttl 300 Cache TTL for resources/list
list_prompts_ttl 300 Cache TTL for prompts/list
read_resource_ttl 3600 Cache TTL for resources/read
get_prompt_ttl 3600 Cache TTL for prompts/get
call_tool_ttl 3600 Cache TTL for tools/call
max_item_size 1048576 Maximum cached item size in bytes (1 MB)
excluded_tools See above Tools that are never cached (mutating or non-deterministic)

Redis Store (Multi-Pod)

Enables Redis-backed session and event storage for multi-replica deployments:

MCP_STORE_CONFIG = {
    "enabled": False,
    "CACHE_REDIS_URL": None,
    "event_store_max_events": 100,
    "event_store_ttl": 3600,
}
Key Default Description
enabled False Enable Redis-backed store
CACHE_REDIS_URL None Redis connection URL (e.g., redis://redis-host:6379/0)
event_store_max_events 100 Maximum events retained per session
event_store_ttl 3600 Event TTL in seconds

Session & CSRF

These values are flat-merged into the Flask app config used by the MCP server process:

MCP_SESSION_CONFIG = {
    "SESSION_COOKIE_HTTPONLY": True,
    "SESSION_COOKIE_SECURE": False,
    "SESSION_COOKIE_SAMESITE": "Lax",
    "SESSION_COOKIE_NAME": "superset_session",
    "PERMANENT_SESSION_LIFETIME": 86400,
}

MCP_CSRF_CONFIG = {
    "WTF_CSRF_ENABLED": True,
    "WTF_CSRF_TIME_LIMIT": None,
}

Troubleshooting

Server won't start

  • Verify fastmcp is installed: pip install fastmcp
  • Check that MCP_DEV_USERNAME is set if auth is disabled -- the server requires a user identity
  • Confirm the port is not already in use: lsof -i :5008

401 Unauthorized

  • Verify your JWT token has not expired (exp claim)
  • Check that MCP_JWT_ISSUER and MCP_JWT_AUDIENCE match the token's iss and aud claims exactly
  • For RS256 with JWKS: confirm the JWKS URI is reachable from the MCP server
  • For RS256 with static key: confirm the public key string includes the BEGIN/END markers
  • For HS256: confirm the secret matches between the token issuer and MCP_JWT_SECRET
  • Enable MCP_JWT_DEBUG_ERRORS = True for detailed server-side logging (errors are never leaked to the client)

Tool not found

  • Ensure the MCP server and Superset share the same superset_config.py
  • Check server logs at startup -- tool registration errors are logged with the tool name and reason

Client can't connect

  • Verify the MCP server URL is reachable from the client machine
  • For Claude Desktop: fully quit the app (not just close the window) and restart after config changes
  • For remote access: ensure your firewall and reverse proxy allow traffic to the MCP port
  • Confirm the URL path ends with /mcp (e.g., http://localhost:5008/mcp)

Permission errors on tool calls

  • The MCP server enforces Superset's RBAC permissions -- the authenticated user must have the required roles
  • In development mode, ensure MCP_DEV_USERNAME maps to a user with appropriate roles (e.g., Admin)
  • Check superset/security/manager.py for the specific permission tuples required by each tool domain (e.g., ("can_execute_sql_query", "SQLLab"))

Response too large

  • If a tool call returns an error about exceeding token limits, the response size guard is blocking an oversized result
  • Reduce page_size or limit parameters, use select_columns to exclude large fields, or add filters to narrow results
  • To adjust the threshold, change token_limit in MCP_RESPONSE_SIZE_CONFIG
  • To disable the guard entirely, set MCP_RESPONSE_SIZE_CONFIG = {"enabled": False}

Security Best Practices

  • Use TLS for all production MCP endpoints -- place the server behind a reverse proxy with HTTPS
  • Enable JWT authentication for any internet-facing deployment
  • RBAC enforcement -- The MCP server respects Superset's role-based access control. Users can only access data their roles permit
  • Secrets management -- Store MCP_JWT_SECRET, database credentials, and API keys in environment variables or a secrets manager, never in config files committed to version control
  • Scoped tokens -- Use MCP_REQUIRED_SCOPES to limit what operations a token can perform
  • Network isolation -- In Kubernetes, restrict MCP pod network policies to only allow traffic from your AI client endpoints
  • Review the Security documentation for additional extension security guidance

Next Steps

  • MCP Integration -- Build custom MCP tools and prompts via Superset extensions
  • Security -- Security best practices for extensions
  • Deployment -- Package and deploy Superset extensions
Reference in New Issue View Git Blame Copy Permalink