mcp-streamable-http-bridge

Bridge stdio Model Context Protocol (MCP) servers to MCP Streamable HTTP behind a single gateway. Each client supplies tenant-specific values via HTTP headers; the wrapper spawns isolated subprocesses, routes MCP traffic, evicts idle instances, and exposes logs and Prometheus metrics.

Objective

Stdio MCP servers assume one long-lived process per client with environment variables fixed at startup. That works for a desktop client (for example Cursor) but not for a shared, multi-user gateway.

This project:

• Loads MCP server definitions from a config folder using the same shape as Cursor mcp.json (mcpServers).

• Accepts HTTP requests and forwards them to the correct stdio child process.

• Maps x-<ENV_NAME> request headers to process environment variables declared under each server’s env block.

• Derives a stable hash from the server name plus resolved env values to reuse one subprocess per tenant.

• Evicts subprocesses after a configurable idle period (global setting).

• Ships with Node.js, npx, uv, and python3 in the container image for common MCP launch patterns.

• Logs with Winston (console + size-rotated file) and exposes Prometheus metrics.

Implementation is Node.js (JavaScript), organized in a class-first structure.

Related MCP server: MCP HTTP-stdio Proxy

Quick start

Docker Compose

Pull the published image from ghcr.io/elad-bar/mcp-streamable-http-bridge (built on push to main via GitHub Actions), or build locally:

docker compose pull
docker compose up -d
# or build locally (tags the same GHCR name):
docker compose up --build -d

• Config: edit JSON under config/ (mounted read-only at /config in the container).

• Logs: written to logs/ on the host via ./logs:/app/logs (active file /app/logs/mcp-wrapper.log inside the container).

• Env: copy .env.example to .env to override defaults; Compose also reads optional .env for PORT on the host mapping.

Health: http://localhost:8080/ (HTML catalog) · Raw Markdown: /catalog.md · Metrics: http://localhost:8080/metrics

Docker (manual)

docker build -t ghcr.io/elad-bar/mcp-streamable-http-bridge:latest .
docker run --rm -p 8080:8080 \
  -v /path/to/mcp-config:/config:ro \
  -v /path/to/logs:/app/logs \
  -e MCP_WRAPPER_IDLE_MINUTES=15 \
  ghcr.io/elad-bar/mcp-streamable-http-bridge:latest

Local development

npm install
npm start    # reads ./config, writes ./logs/mcp-wrapper.log
npm test

Outside Docker, paths resolve to the repo config/ and logs/ directories. In the container they stay /config and /app/logs/mcp-wrapper.log (via /.dockerenv). Integration tests override paths via constructor options.

Configuration

Gateway config (config/)

The wrapper reads only config/mcp.json for gateway mcpServers. Other files in config/ (e.g. SSH MCP ssh-config.json, keys under config/keys/) are for child processes only and are not parsed by the gateway.

Each loaded file uses the gateway schema (stdio spawn line):

{
  "mcpServers": {
    "grafana": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-grafana"],
      "env": {
        "GRAFANA_URL": "https://grafana.example.com",
        "GRAFANA_API_KEY": ""
      }
    }
  }
}

• command / args: spawn line for the MCP server (stdio).

• env: keys injected into the child process. Values may be defaults in JSON and/or overridden by headers (see below).

Every key listed under env participates in tenant identity (hash) and in the subprocess environment.

Client config (Streamable HTTP)

MCP clients (e.g. Cursor) connect to the gateway, not to stdio directly. Use url pointing at /mcp/{serverName} and optional headers for tenant x-<ENV> values (same names as in the gateway env block):

{
  "mcpServers": {
    "grafana": {
      "url": "http://localhost:8080/mcp/grafana",
      "headers": {
        "x-GRAFANA_URL": "https://grafana.example.com",
        "x-GRAFANA_API_KEY": "<your-api-key>"
      }
    }
  }
}

The client performs MCP Streamable HTTP (initialize, session id, tool calls). You do not configure initialize by hand—only the gateway URL and tenant headers.

Open GET / in a browser for a styled catalog with per-server client snippets, or GET /catalog.md / Accept: text/markdown for raw Markdown. Set PUBLIC_BASE_URL when the gateway is behind a reverse proxy so URLs in the catalog use the public host.

HTTP usage (MCP Streamable HTTP)

The wrapper implements MCP Streamable HTTP on /mcp/{serverName}:

1. First request — POST with method: "initialize", tenant x-* headers, and JSON-RPC body. Response includes mcp-session-id header.

2. Later requests — POST with the same mcp-session-id and one JSON-RPC message per request (no tenant headers required if session is valid).

3. Server → client streaming — GET with Accept: text/event-stream and mcp-session-id to receive MCP notifications over SSE.

Tenant headers

For each key in the server’s env object, send a header:

x-<ENV_NAME>: <value>

Header names are case-insensitive. Examples:

Resolution order: header value if present, otherwise the value from JSON. If a key has no value from either source, the request is rejected before spawning.

Instance key: sha256(serverName + canonical resolved env) — same server + same env → same subprocess; any difference → a new subprocess.

Do not log header values or full instance keys; use structured logs with a short hash prefix for support.

Stdio children use Content-Length framed JSON-RPC on stdin/stdout (same as the official MCP SDK).

Environment variables

Gateway

Logging (Winston)

Fixed paths (not env-configurable): In Docker, MCP JSON under /config, logs /app/logs/mcp-wrapper.log, Prometheus at GET /metrics. Locally, config/ and logs/mcp-wrapper.log under the repo root. Bind-mount host config/ and logs/ in Compose.

Metrics (Prometheus)

Aggregated metrics avoid high-cardinality labels (no full tenant hash on time series). Typical series:

• mcp_servers_configured — count after config load

• mcp_instances_active{server} — running subprocesses per configured server

• mcp_rpc_requests_total{server,mcp_method} — MCP methods (tools/list, tools/call, …)

• mcp_tool_calls_total{server,tool} — tools/call by tool name

• mcp_rpc_duration_seconds{server,mcp_method} — latency histograms

Per-tenant detail belongs in structured logs, not metric labels. See docs/architecture.md.

Documentation

License

Add your license file as appropriate for your distribution.