Running Multiple MCP Servers on Docker + Windsurf and Claude

MCP Notes – Working Draft

I’m sharing my raw notes on MCP servers I captured between family time and hacking over the past few days. They’re rough but runnable and might save you the same rabbit‑holes I fell into. Includes information on LAN sharing, proxies, bridging, and https://smithery.ai.

MCP Server Types

Two main types of MCP servers, differentiated by transport:

Local / stdio – runs on the same machine; easiest via Docker.
Remote / SSE – reachable over HTTP(S); can be LAN or public.

supergateway turns a local stdio server into an SSE endpoint so every client can connect.

Finding & Running MCP Servers

Server	Where to get it	Why it matters
GitHub MCP	github-mcp-server	Everybody has a GitHub account; perfect “hello-world” server.
Context7 MCP	`@upstash/context7-mcp` (npm)	Great for natural‑language context queries.
Filesystem MCP	filesystem server	Lets an AI traverse a directory as context.
Registry / search	https://smithery.ai	Directory + installer for everything MCP‑related.

Quick Docker commands

1. GitHub MCP – local stdio test
Runs entirely inside the container; no proxies needed.

docker run -i --rm \
  -e GITHUB_PERSONAL_ACCESS_TOKEN=your_token_here \
  -p 8000:8000 \
  ghcr.io/github/github-mcp-server:latest

2. Context7 MCP via supergateway
Exposes Context7 over /sse so clients can fetch http://localhost:8010/sse or http://server-ip:8010/sse. For LAN clients with Windsurf, see note on socat workaround.

docker run -i --rm -p 8010:8000 supercorp/supergateway \
  --stdio "npx -y @upstash/context7-mcp" \
  --port 8000 \
  --ssePath /sse \
  --messagePath /message

3. Filesystem MCP via supergateway
Mounts your current directory read‑only for the AI to browse.

docker run -i --rm -p 8011:8000 \
  -v $(pwd):/projects:ro \
  supercorp/supergateway \
  --stdio "npx -y @modelcontextprotocol/server-filesystem /projects" \
  --port 8000 \
  --ssePath /sse \
  --messagePath /message

(Swap the host‑side ports if 8010/8011 collide with something else.)

Token reminder – For anything beyond a throw‑away test, load secrets from a .env file or --env-file. Avoid hard‑coding tokens inline.

Windsurf: auto‑run vs client‑only config

Auto‑run entry (spawns container on demand)

"github-local": {
  "command": "docker",
  "args": [
    "run","-i","--rm",
    "-e","GITHUB_PERSONAL_ACCESS_TOKEN",
    "-p","9000:8000",
    "ghcr.io/github/github-mcp-server:latest"
  ],
  "env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "your-token" }
}

Windsurf launches this when you pick the server in the UI. Endpoint becomes http://localhost:9000/sse.

Plain client configuration (remote server)

{
  "mcpServers": {
    "context7-remote": {
      "serverUrl": "http://server-ip:8010/sse"
    }
  }
}

Use this for a remote or LAN‑hosted server. On macOS you’ll need the socat workaround below.

Sandbox issue & socat workaround

Windsurf’s MacOS build ships with a hardened network sandbox: only localhost or external addresses are allowed. If your MCP runs on a LAN box (or even a LAN IP on the same box when Docker uses host‑network), you’ll see “no route to host”.

socat TCP-LISTEN:9000,bind=127.0.0.1,reuseaddr,fork \
      TCP:192.168.1.95:8010

Then point Windsurf at http://localhost:9000/sse. Linux users should be able to launch Windsurf with --no-sandbox, though this is not recommended practice.

Transport Helpers

Tool	Direction	Core role	Typical use case
supergateway	→	Wrap stdio server → expose as SSE	Share a local Docker container with any SSE‑capable client.
mcp‑proxy	↔	Reverse‑proxy / bridge an SSE server to a new local endpoint	Let Claude Desktop reach a LAN or remote server through `localhost`.
~~mcp‑remote~~	←	Legacy: stdio‑only client → remote SSE server	Rare; shell scripts that still speak stdio.

Example: mcp‑proxy entry in Claude Desktop

First install mcp-proxy. Easiest ways are either:

uv tool install mcp-proxy or
npx -y @smithery/cli install mcp-proxy --client claude.

You will probably need to specify the full path to mcp-proxy as shown below, especially if installed via uv.

"context7-proxy": {
  "command": "/Users/manta/.local/bin/mcp-proxy",
  "args": [ "http://192.168.1.95:8010/sse" ]
}

View settings in Claude Desktop > Settings > Integrations. Edit directly in claude-desktop-config.json.

Claude now connects via the proxy. Windsurf still needs the socat workaround if the server lives off‑localhost.

Smithery in one minute

Smithery.ai is a unified registry and installer for Model Context Protocol (MCP) servers and plugins, making it easy to extend AI agents like Claude or ChatGPT with new tools and integrations.

Centralized Discovery & Install: Browse, install, and manage MCP-compatible servers and plugins with a single CLI command.
Seamless Integration: Instantly connects AI clients to new APIs, automations, and data sources—locally or in the cloud—with secure config handling.
Standardized & Community-Driven: Ensures all integrations follow open standards and are easy to maintain or share.

Get started at https://smithery.ai

Still on the backlog

Wrap MCP with ChatGPT tool calling for use inOpenAI desktop app.
Better error handling when GitHub MCP hits rate limits or drops SSE.

← Back to Home

Erik Corkran — Blog