mcpd

One host for all your MCP servers.

Every MCP client — Claude Code, Claude Desktop, Cursor — spawns its own copy of every server you've configured, per window. Dozens of invisible processes, gigabytes of RAM, no way to see any of it.

mcpd is a single daemon that owns your servers instead:

• One shared instance per server, across every client and window
• Lazy start — servers spawn on first call, stop after 5 idle minutes
• Visibility — live dashboard and mcpd status: state, memory, calls, last used
• Zero dependencies — plain Node 20+

$ mcpd status
daemon pid 75164 · port 11431 · 6 servers
  filesystem           running pid 75186 · 74.4MB   calls 1     last used 51s ago
  memory               running pid 75272 · 73.0MB   calls 6     last used 3s ago
  sequential-thinking  idle                         calls 0     last used never
  github               running pid 75190 · 67.7MB   calls 1     last used 17s ago
  total server memory: 460.0MB

Bonus: sessions share server state — store a fact in the memory server from one Claude session, recall it from another.

Quick start

$ git clone https://github.com/milind-soni/mcpd && cd mcpd && npm link
$ mcpd import claude-code        # or claude-desktop / cursor
$ mcpd start

Dashboard: http://127.0.0.1:11431/

Point your clients at mcpd instead of spawning servers:

// HTTP clients (Claude Code, Cursor, ...)
{ "type": "http", "url": "http://127.0.0.1:11431/servers/<name>/mcp" }

// stdio-only clients
{ "command": "mcpd", "args": ["connect", "<name>"] }

Add a server:

$ mcpd add github -- npx -y @modelcontextprotocol/server-github

How it works

Claude Code ──┐
Cursor ───────┼── HTTP ──►  mcpd daemon  ──┬── filesystem (shared)
more windows ─┘             :11431         ├── github     (shared)
                                           └── memory     (shared)

mcpd speaks streamable HTTP to clients and stdio to the servers it supervises. Each server's initialize handshake happens once; every client's handshake is answered from cache, and requests are forwarded with remapped ids. To each client it looks like a private server.

Native servers

mcpd supervises any command — including MCP servers compiled to native binaries with scriptc.

examples/sequential-thinking-native is Anthropic's official sequential-thinking server ported to 259 lines of plain TypeScript and compiled statically. Tool outputs byte-identical to the original:

	native (scriptc)	official (node)
size on disk	554KB	~102MB (package + Node)
startup	1.9ms	86ms
memory	1.6MB	79MB

45 copies of the binary use less RAM than one copy of the Node original.

Commands

mcpd start [--foreground]   start the daemon
mcpd stop                   stop the daemon and all servers
mcpd status                 servers, memory, calls, last used
mcpd add <name> -- <cmd>    register a server
mcpd remove <name>          unregister a server
mcpd url <name>             print a server's HTTP endpoint
mcpd connect <name>         stdio bridge for legacy clients
mcpd import <source>        claude-code | claude-desktop | cursor

Config: ~/.mcpd/config.json · Logs: ~/.mcpd/mcpd.log

Limitations (v0)

• Server-initiated messages (sampling, notifications) are not routed — tools/resources/prompts work fine
• Config changes need a restart
• Shared-by-default; per-server isolate mode is planned
• macOS/Linux

Development

$ npm test

MIT © Milind Soni