mcpField
M
mcpgo

mcpgo

Local
github.com/jschoemakerDeveloper Tools0Updated 23w ago

About mcpgo

mcpgo is an MCP server for managing MCP configurations in Claude Code through conversational commands. It can wrap stdio servers for automatic process restarts and also supports Codex CLI MCPs.

Key features

  1. 01List and inspect configured MCPs
  2. 02Add, remove, and configure servers
  3. 03Wrap stdio MCPs for restarts
  4. 04Restart wrapped server processes
  5. 05Check server process health
  6. 06Manage Codex CLI MCPs

Use cases

  1. 01List MCPs while using Remote Control
  2. 02Restart a crashed or unresponsive wrapped server
  3. 03Adjust an MCP server configuration through Claude

MCP support

mcpgo is itself an MCP server, exposing tools to manage Claude Code MCP configurations and processes. It can also wrap and restart stdio MCPs, and includes tools for Codex CLI MCPs.

Configuration

Local server config is the same for every client. Copy it into your client’s MCP config file.

mcp.json · Universal config
{
"mcpServers": {
"jschoemaker-mcpgo": {
"command": "npx",
"args": [
"-y",
"@exfil/mcpgo"
]
}
}
}
Run commandnpx -y @exfil/mcpgo

Local servers need a runtime on your machine (e.g. Node.js 18+). Config file locations differ by client; see your client’s docs.

Packages
npm@exfil/mcpgo0.1.6

Tools

11

Saved tool list · updated 3w ago

ParameterTypeRequiredDescription
show_env_valuesbooleanNoIf true, reveal raw 'env' values. Default false (values shown as '***').

Details

From the project README

mcpgo

Manage your Claude Code MCPs by talking to Claude — list, restart, wrap, and more.

Works everywhere Claude Code works, including Remote Control where /mcp isn't available.

Installation

npm install -g @exfil/mcpgo
claude mcp add mcpgo --scope user -- mcpgo

Requirements: Node.js ≥ 18

Remote Control

Claude Code's Remote Control mode has no access to /mcp. That means you can't list, restart, or manage MCP servers remotely — unless you have mcpgo.

Because mcpgo is itself an MCP server, it's available everywhere Claude Code is:

"list my mcps"
"restart github"
"check postgres health"

No /mcp UI needed.

What it does

Reliable restarts

Claude Code connects to stdio MCP servers via pipes it creates at launch. If a server crashes or you want to restart it, Claude Code marks it as failed with no auto-recovery.

mcpgo solves this with a wrapper: a long-lived process that sits between Claude Code and your MCP. Claude Code talks to the wrapper (which stays alive), and the wrapper spawns the real server as a child — auto-respawning it whenever it's killed.

Wrapping is optional — you can use mcpgo just for listing, adding, or removing MCPs without wrapping anything. Wrapping is only needed if you want reliable restarts without touching /mcp.

One-time setup per MCP:

  1. wrap_mcp_stdio — wraps the target MCP in config
  2. Restart Claude Code (or disconnect + reconnect via /mcp)
  3. Done — use restart_mcp_process anytime

After setup:

"restart github"     → instant, wrapper auto-respawns the child, Claude Code stays connected
"restart postgres"        → same

If you ever want to undo wrapping, use unwrap_mcp_stdio to restore the original config.

Tools

Process management
wrap_mcp_stdio

Wrap a Claude Code stdio MCP for reliable restarts.

"wrap github"
"wrap postgres"
unwrap_mcp_stdio

Restore a wrapped MCP to its original config.

"unwrap github"
"unwrap postgres"
restart_mcp_process

Restart a wrapped MCP's child process — wrapper auto-respawns it, Claude Code stays connected.

"restart github"
"restart postgres"
check_mcp_health

Check if an MCP is configured, wrapped, and its process is running.

"check github health"
"is postgres running?"
Codex CLI support
wrap_codex_mcp_stdio

Wrap a Codex CLI MCP from ~/.codex/config.toml.

"wrap my codex mcp called postgres"
restart_codex_mcp_process

Restart a wrapped Codex CLI MCP.

"restart the codex postgres"
Config management
list_all_mcps

List all configured MCP servers.

"list my mcps"
"what mcps do I have?"
add_mcp

Add a new MCP server to ~/.claude.json.

"add an mcp called my-tool that runs python server.py"
remove_mcp

Remove an MCP server.

"remove the my-tool mcp"
configure_mcp

Update an existing MCP server's config fields.

"change github's command to python3"
get_mcp_details

Get full config details for a specific MCP.

"show me the github config"
"get details for postgres"

How wrapping works

When you call wrap_mcp_stdio "github", the config entry changes from:

{ "command": "python", "args": ["server.py"] }

to:

{
  "command": "node",
  "args": [
    "/path/to/wrapper.js",
    "--name", "github",
    "--pidfile", "...",
    "--env-allowlist", "PATH,Path,PATHEXT,HOME,USERPROFILE,SystemRoot,...",
    "--",
    "python", "server.py"
  ]
}

The wrapper:

  • Spawns the original command as a child
  • Writes the child's PID to a pidfile
  • Auto-respawns the child on exit (with backoff)
  • Passes a curated env to the child (PATH, locale, temp, home, plus the MCP's declared env keys) instead of inheriting Claude Code's full environment
  • Exits cleanly on SIGTERM, SIGINT, or stdin EOF

restart_mcp_process reads the pidfile and kills the child. The wrapper respawns it — Claude Code never sees a disconnect.

unwrap_mcp_stdio reverses this — it restores the original command from the args after -- and removes the pidfile.

Stable wrapper location

When you wrap an MCP, mcpgo copies wrapper.js to a stable user-data location and writes that path into ~/.claude.json. This means the config survives npx cache clears and mcpgo version upgrades — the wrapper won't disappear from under you.

  • Windows: %LOCALAPPDATA%\mcpgo\wrapper.js
  • Linux/macOS: ~/.mcpgo/wrapper.js

Re-running wrap_mcp_stdio on an already-wrapped MCP refreshes the wrapper copy in place.

Pidfile locations

  • Windows: %LOCALAPPDATA%\mcpgo\pids\<name>.pid
  • Linux/macOS: $TMPDIR/mcpgo/pids/<name>.pid

Notes

  • Wrapping requires a Claude Code restart (or manual /mcp disconnect + reconnect) to take effect — Claude Code caches config in memory
  • restart_mcp_process works best on wrapped MCPs; for unwrapped ones it kills the process but you'll need to reconnect manually via /mcp
  • Mid-request crashes may cause a brief protocol desync; Claude Code recovers on the next tool call
  • Wrapped MCPs receive a curated env (PATH, locale, temp, home, plus the MCP's declared env keys) rather than the full Claude Code environment. Re-run wrap_mcp_stdio on previously-wrapped MCPs to pick up this default
  • list_all_mcps and get_mcp_details redact env values by default so the tools can't be used to exfiltrate secrets stored in MCP configs — pass show_env_values: true when you actually need to see them

Comments

Markdown supported

New to MCP?

From what the protocol is to installing your first server, these answers cover most first questions.

An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.