MCP 服务器

bonnard

Open-source agentic schema layer. Define metrics once in YAML, query governed data from any warehouse (Snowflake, BigQuery, Databricks, PostgreSQL, DuckDB) via MCP.

README

Self-hosted semantic layer for AI agents.

<a href="https://docs.bonnard.dev/docs/">Docs</a> · <a href="https://www.npmjs.com/package/@bonnard/cli">CLI</a> · <a href="https://discord.com/invite/RQuvjGRz">Discord</a> · <a href="https://www.bonnard.dev">Website</a>

Bonnard is an agent-native semantic layer — one set of metric definitions, every consumer (AI agents, apps, dashboards) gets the same governed answer. This repo is the self-hosted Docker deployment: run Bonnard on your own infrastructure with no cloud account needed.

Quick Start

# 1. Scaffold project
npx @bonnard/cli init --self-hosted

# 2. Configure your data source
#    Edit .env with your database credentials

# 3. Start the server
docker compose up -d

# 4. Define your semantic layer
#    Add cube/view YAML files to bonnard/cubes/ and bonnard/views/

# 5. Deploy models to the server
bon deploy

# 6. Verify your semantic layer
bon schema

# 7. Connect AI agents
bon mcp

Requires Node.js 20+ and Docker.

What's Included

MCP server — AI agents query your semantic layer over the Model Context Protocol
Cube semantic layer — SQL-based metric definitions with caching, access control, and multi-database support
Cube Store — pre-aggregation cache for fast analytical queries
Admin UI — browse deployed models, views, and measures at http://localhost:3000
Deploy API — push model updates via bon deploy without restarting containers
Health endpoint — GET /health for uptime monitoring

Connecting AI Agents

Run bon mcp to see connection config for your setup. Examples below.

Claude Desktop / Cursor

{
  "mcpServers": {
    "bonnard": {
      "url": "https://bonnard.example.com/mcp",
      "headers": {
        "Authorization": "Bearer your-secret-token-here"
      }
    }
  }
}

Claude Code

{
  "mcpServers": {
    "bonnard": {
      "type": "url",
      "url": "https://bonnard.example.com/mcp",
      "headers": {
        "Authorization": "Bearer your-secret-token-here"
      }
    }
  }
}

CrewAI (Python)

from crewai import MCPServerAdapter

mcp = MCPServerAdapter(
    url="https://bonnard.example.com/mcp",
    transport="streamable-http",
    headers={"Authorization": "Bearer your-secret-token-here"}
)

Production Deployment

Authentication

Protect your endpoints by setting ADMIN_TOKEN in .env:

ADMIN_TOKEN=your-secret-token-here

All API and MCP endpoints will require Authorization: Bearer <token>. The /health endpoint remains open for monitoring.

Restart after changing .env:

docker compose up -d

TLS with Caddy

Caddy provides automatic HTTPS via Let's Encrypt.

Create a Caddyfile next to your docker-compose.yml:

bonnard.example.com {
    reverse_proxy localhost:3000
}

Add Caddy to your docker-compose.yml:

  caddy:
    image: caddy:2
    ports:
      - "80:80"
      - "443:443"
    volumes:
      - ./Caddyfile:/etc/caddy/Caddyfile:ro
      - caddy_data:/data
    restart: unless-stopped

Add the volume at the top level:

volumes:
  models: {}
  caddy_data: {}

Then remove the Bonnard port mapping (ports: - "3000:3000") since Caddy handles external traffic.

Deploy to a VM

# Copy project files to your server
scp -r . user@your-server:~/bonnard/

# SSH in and start
ssh user@your-server
cd ~/bonnard
docker compose up -d

Configuration

Variable	Description	Default
`CUBEJS_DB_TYPE`	Database driver (`postgres`, `duckdb`, `snowflake`, `bigquery`, `databricks`, `redshift`, `clickhouse`)	`duckdb`
`CUBEJS_DB_*`	Database connection settings (host, port, name, user, pass)	—
`CUBEJS_DATASOURCES`	Comma-separated list for multi-datasource setups	`default`
`CUBEJS_API_SECRET`	HS256 secret for Cube JWT auth (auto-generated by `bon init`)	—
`ADMIN_TOKEN`	Bearer token for API/MCP authentication	— (open)
`CUBE_PORT`	Cube API port	`4000`
`BONNARD_PORT`	Bonnard server port	`3000`
`CORS_ORIGIN`	Allowed CORS origins	`*`
`CUBE_VERSION`	Cube Docker image tag	`v1.6`
`BONNARD_VERSION`	Bonnard Docker image tag	`latest`

See .env.example for a full annotated configuration file.

Architecture

Service	Image	Role
`cube`	`cubejs/cube`	Semantic layer engine — executes queries against your warehouse
`cubestore`	`cubejs/cubestore`	Pre-aggregation cache — stores materialized results for fast reads
`bonnard`	`ghcr.io/bonnard-data/bonnard`	MCP server, admin UI, deploy API — the interface layer for agents and tools

All three services communicate over an internal Docker network. Only bonnard (port 3000) and optionally cube (port 4000) are exposed externally.

Monitoring

# Health check
curl http://localhost:3000/health

# View logs
docker compose logs -f

# View active MCP sessions
curl -H "Authorization: Bearer <token>" http://localhost:3000/api/mcp/sessions

Deploying Schema Updates

From your development machine:

bon deploy

This pushes your cube/view YAML files to the running server. No restart needed — Cube picks up changes automatically.

Pinning Versions

Control image versions via .env:

CUBE_VERSION=v1.6
BONNARD_VERSION=latest

Supported Data Sources

Warehouses: Snowflake, Google BigQuery, Databricks, PostgreSQL (including Supabase, Neon, RDS), Amazon Redshift, DuckDB (including MotherDuck), ClickHouse

See the full documentation for connection guides.