mcp-local-hub

Run one copy of each Model Context Protocol server on your workstation, shared across every MCP client that needs it — instead of each client spawning its own redundant stdio process.

The problem

Every modern coding assistant (Claude Code, Codex CLI, Gemini CLI, Antigravity, Cursor, Continue, …) speaks MCP, and each client independently execs whatever stdio servers you configure — uvx serena, npx @modelcontextprotocol/server-memory, mcp-language-server, and so on. If you use three assistants side-by-side on the same project, you get three Serena processes, three gopls subprocesses, three separate memory stores. Each per-session spawn re-downloads dependencies, re-indexes your code, and competes for RAM.

What this tool does

mcp-local-hub runs each MCP server once per OS user, exposes it as a local HTTP endpoint via Streamable HTTP transport, and writes the correct client-config entry into each managed MCP client. Clients see a shared daemon instead of their own child process.

   ┌─────────────────────────────────────────────────────────────────────┐
   │             OS-level Task Scheduler (Windows schtasks)              │
   │             starts on logon, restarts on failure                    │
   └──┬──────┬──────┬──────┬──────┬──────┬──────┬──────┬─────────────────┘
      │      │      │      │      │      │      │      │
      ▼      ▼      ▼      ▼      ▼      ▼      ▼      ▼
   ┌─────┐┌─────┐┌─────┐┌─────┐┌─────┐┌─────┐┌─────┐┌─────┐
   │seren││seren││memor││seq- ││wolf-││god- ││paper││time │
   │claud││codex││y    ││think││ram  ││bolt ││-srch││     │
   │9121 ││9122 ││9123 ││9124 ││9125 ││9126 ││9127 ││9128 │
   └──┬──┘└──┬──┘└──┬──┘└──┬──┘└──┬──┘└──┬──┘└──┬──┘└──┬──┘
      │      │      │      │      │      │      │      │
      │      │      └──────┴──────┴──────┴──────┴──────┴── (shared by all 4 clients)
      │      │
   ┌──┼──────┼──────────┬──────────┐
   ▼  ▼      ▼          ▼          ▼
  Claude   Gemini    Antigravity  Codex CLI
  Code     CLI       (stdio       (HTTP)
  (HTTP)   (HTTP)     relay)

The 6 non-serena daemons run as stdio-bridge via a native Go stdio-host (internal/daemon/host.go): one subprocess per daemon, multiplexed across concurrent HTTP clients via JSON-RPC id rewriting and a cached initialize response.

Antigravity's Cascade agent rejects loopback-HTTP MCP entries, so mcp-local-hub bridges it via a stdio relay subprocess: mcphub.exe relay translates between stdio JSON-RPC and the shared HTTP daemon. Cascade sees a normal stdio command; the daemon stays shared.

Quick start

# Build
go build -o mcphub.exe ./cmd/mcphub

# Install all bindings (2 daemons + 4 client configs)
./mcphub.exe install --server serena

# Verify
./mcphub.exe status
claude mcp get serena    # shows: Status: ✓ Connected, Type: http

Detailed setup, per-client behaviour, and troubleshooting in INSTALL.md.

Supported clients

| Client | Version tested | Config path | Transport | |---|---|---|---| | Claude Code CLI | 2.1.112 | ~/.claude.json | HTTP (type: "http") | | Codex CLI | 0.121.0 | ~/.codex/config.toml | HTTP (streamable_http) | | Gemini CLI | 0.38.1 | ~/.gemini/settings.json | HTTP (type: "http") | | Antigravity IDE | v0.x | ~/.gemini/antigravity/mcp_config.json | stdio relay → HTTP |

Antigravity note: Cascade rejects loopback-HTTP MCP entries, so mcp-local-hub writes a stdio relay entry instead — mcphub.exe relay --server serena --daemon claude. Cascade spawns the relay as a normal stdio subprocess; the relay forwards JSON-RPC to the shared HTTP daemon on port 9121. No extra Serena process per Antigravity session.

Key commands

| Command | What it does | |---|---| | mcphub install --server <name> | Create scheduler tasks for each daemon, write client config entries, start daemons | | mcphub install --server <name> --daemon <d> | Install only one daemon + its referencing client bindings | | mcphub install --server <name> --dry-run | Print the plan without applying | | mcphub uninstall --server <name> | Reverse: delete scheduler tasks, remove client entries (backups retained) | | mcphub rollback | Restore the latest .bak-mcp-local-hub-* for every client | | mcphub status | Show state of all mcp-local-hub-* Task Scheduler tasks | | mcphub restart --server <name> \| --all | Stop + re-run scheduler tasks | | mcphub daemon --server <n> --daemon <d> | Invoked by the scheduler; exec the real server with tee'd logs | | mcphub relay --server <n> --daemon <d> | stdio↔HTTP bridge for clients that reject loopback HTTP (e.g. Antigravity) | | mcphub relay --url <url> | Direct relay to an arbitrary Streamable HTTP endpoint | | mcphub secrets {init,set,get,list,delete,edit,migrate} | Manage age-encrypted vault for API keys etc. | | mcphub version | Print build version, commit, and date |

Current status

Phase 2 complete (2026-04-17). 8 MCP daemons consolidated:

serena (×2 contexts: claude-code, codex)
memory, sequential-thinking, wolfram, godbolt, paper-search-mcp, time

Plus context7 as direct HTTPS entry (no daemon needed). Native Go stdio-host (internal/daemon/host.go) replaces supergateway npm dep for stdio→HTTP bridging. All 4 clients (Claude Code, Codex CLI, Gemini CLI, Antigravity) share these daemons. See docs/phase-2-verification.md.

Roadmap (not yet implemented):

Phase 3: workspace-scoped daemons + mcp register/unregister for per-project mcp-language-server instances
Post-Phase-2: minimal GUI installer (scans client configs, checkboxes for hub routing)
Phase 4+: Linux/macOS scheduler backends

Platform support

Windows 11 is first-class (tested on 10.0.26100). Linux and macOS ship compile-only stubs in internal/scheduler/ — the build succeeds on GOOS=linux and GOOS=darwin, but mcphub install fails immediately with "not yet implemented" on those platforms. Real systemd-user-unit and launchd-agent backends are Phase 4 scope.

License

Apache License 2.0 — see LICENSE.

MCP Servers