MCP Gateway
Aggregates multiple upstream MCP servers (stdio, streamable-http, sse, websocket) behind a single streamable-HTTP endpoint with declarative configuration, namespacing, filtering, and auto-reload.
README
MCP Gateway
A FastMCP 2.x-based gateway that aggregates multiple
upstream Model Context Protocol servers
(stdio, streamable-http, sse, websocket) behind a single streamable-HTTP
endpoint. Upstream servers are configured declaratively in
config/mcpServers.json, with per-server namespacing, tool/resource/prompt
filtering, read-only enforcement, environment-variable interpolation, and
automatic reload on config changes.
Architecture
gateway/
├── main.py # entrypoint: builds & serves the gateway, wires reload
├── config.py # GatewayConfig, loaded from environment variables
├── loader.py # mcpServers.json parsing, env interpolation, validation
├── proxy.py # builds the aggregated FastMCP server (mount per upstream)
├── namespace.py # per-server middleware: toolFilter/resourceFilter/promptFilter, readOnly
├── reconnect.py # retry-with-backoff helper used for upstream connectivity probes
├── reload.py # watchdog-based watcher for mcpServers.json
├── health.py # /health, /healthz, /metrics routes
├── logging.py # logging setup
└── transports/
├── stdio.py # spawns local processes (npx, uvx, ...)
├── streamable_http.py # streamable-HTTP upstream client transport
├── sse.py # SSE upstream client transport
└── websocket.py # WebSocket upstream client transport (custom, mcp SDK based)
Each enabled upstream server becomes a FastMCP proxy sub-server
(FastMCPProxy with an explicit client_factory that builds a fresh
transport per call, for concurrent-request isolation), gets a
ServerFilterMiddleware attached
(enforcing readOnly / toolFilter / resourceFilter / promptFilter), and
is mounted onto the main gateway server with namespace=<namespace> when
prefixTools is enabled. Tool/resource/prompt names are then automatically
namespaced by FastMCP as {namespace}_{name}.
Configuration
config/mcpServers.json
{
"mcpServers": {
"filesystem": {
"type": "stdio",
"enabled": true,
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "${FS_ROOT:-/data}"],
"env": {},
"namespace": "fs",
"readOnly": false,
"prefixTools": true,
"toolFilter": ["*"],
"resourceFilter": ["*"],
"promptFilter": ["*"],
"timeout": 30
}
}
}
Field reference per server entry:
| Field | Type | Description |
|---|---|---|
type |
stdio | streamable-http | sse | websocket |
Upstream transport kind |
enabled |
bool | If false, the server is skipped entirely |
url |
string | Required for streamable-http / sse / websocket |
command, args, env |
- | Required for stdio (e.g. npx, uvx) |
headers |
object | Extra HTTP headers for streamable-http / sse |
namespace |
string | Prefix used for this server's tools/resources/prompts; defaults to the server's key |
readOnly |
bool | If true, no tools from this server are registered |
prefixTools |
bool | If true, tools become {namespace}_{tool} |
toolFilter / resourceFilter / promptFilter |
list of glob patterns | fnmatch-style; ["*"] = allow all |
timeout |
number | Per-call timeout (seconds) for the upstream connection probe |
Any string value in the file supports ${VAR} / ${VAR:-default}
interpolation, resolved against the process environment (.env +
docker-compose.yml environment:). An unresolved variable without a
default raises a configuration error for enabled servers; disabled server
entries with unresolved variables are tolerated (logged as a warning).
Environment variables (.env / docker-compose.yml)
| Variable | Default | Description |
|---|---|---|
GATEWAY_LISTEN |
0.0.0.0 |
Bind address (0.0.0.0, 127.0.0.1, localhost supported) |
GATEWAY_PORT |
5555 |
Bind port |
GATEWAY_PATH |
/mcp |
HTTP path for the MCP streamable-HTTP endpoint |
GATEWAY_RELOAD |
true |
Watch mcpServers.json and restart the process on change |
GATEWAY_HEALTH_ENDPOINT |
true |
Enable /health and /healthz |
GATEWAY_METRICS |
true |
Enable /metrics (Prometheus text format) |
GATEWAY_LOG_LEVEL |
info |
debug | info | warning | error | critical |
GATEWAY_PREFIX_TOOLS |
true |
Default for prefixTools when a server entry omits it |
MCP_SERVERS_CONFIG |
/app/config/mcpServers.json |
Path to the servers config file |
Running locally
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
cp .env .env.local # adjust as needed
export $(grep -v '^#' .env.local | xargs)
python -m gateway.main
The gateway will be reachable at http://localhost:5555/mcp, with
http://localhost:5555/health and http://localhost:5555/metrics alongside it.
stdio upstream servers
stdio servers (npx, uvx, ...) are spawned as subprocesses inside the
gateway process/container — no separate container or network hop is needed.
Make sure node/npx and/or uv/uvx are available in the runtime
environment (the provided Dockerfile installs both).
Docker Compose
docker compose up -d --build
This starts only the gateway service, on the mcp Docker network, mounting
./config read-only and loading ./.env. To use the published image instead
of building locally:
docker pull ghcr.io/cociweb/mcp-gateway:latest
docker compose up -d
Reload behavior
When GATEWAY_RELOAD=true, a watchdog observer watches the directory
containing mcpServers.json. On any change to that file, the gateway
gracefully restarts itself in-place (os.execv) to rebuild the full mounted
server tree (reconnect upstreams, reapply filters/prefixes). Streamable-HTTP
clients are expected to reconnect automatically on disconnect.
Health & metrics
GET /health— JSON status, per-server connectivity and config summary.GET /healthz— plain-textokliveness probe.GET /metrics— Prometheus-style text exposition (uptime, enabled/connected server counts, per-server connectivity gauge).
Development / testing
pip install -e ".[dev]"
pytest -q
tests/test_proxy_integration.py spins up a real npx-based MCP filesystem
server and is skipped automatically if npx is not available.
CI / GHCR
.github/workflows/ghcr.yml builds a multi-platform (linux/amd64,
linux/arm64) image and pushes it to ghcr.io/cociweb/mcp-gateway on pushes
to main and on version tags.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。