mcpgo
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@mcpgo list my mcps"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
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 -- mcpgoRequirements: Node.js ≥ 18
Related MCP server: MCP Server Generator
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:
wrap_mcp_stdio— wraps the target MCP in configRestart Claude Code (or disconnect + reconnect via
/mcp)Done — use
restart_mcp_processanytime
After setup:
"restart github" → instant, wrapper auto-respawns the child, Claude Code stays connected
"restart postgres" → sameIf 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
envkeys) instead of inheriting Claude Code's full environmentExits 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.jsLinux/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>.pidLinux/macOS:
$TMPDIR/mcpgo/pids/<name>.pid
Notes
Wrapping requires a Claude Code restart (or manual
/mcpdisconnect + reconnect) to take effect — Claude Code caches config in memoryrestart_mcp_processworks best on wrapped MCPs; for unwrapped ones it kills the process but you'll need to reconnect manually via/mcpMid-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
envkeys) rather than the full Claude Code environment. Re-runwrap_mcp_stdioon previously-wrapped MCPs to pick up this defaultlist_all_mcpsandget_mcp_detailsredactenvvalues by default so the tools can't be used to exfiltrate secrets stored in MCP configs — passshow_env_values: truewhen you actually need to see them
Available Tools
11 toolsadd_mcpB
Add a new MCP server to the user configuration (~/.claude.json).
Inputs: mcp_name (required), transport (required), env, url, +6 more.
check_mcp_healthA
Check the health of an MCP server — whether it is configured, wrapped, and its process is running.
Inputs: mcp_name (required).
configure_mcpC
Update the configuration of an existing MCP server. Merges provided updates into the existing config.
Inputs: updates (required), mcp_name (required), allow_footgun_env, allow_footgun_args, +1 more.
get_mcp_detailsB
Get the full configuration for a specific MCP server by name.
Inputs: mcp_name (required), show_env_values.
list_all_mcpsA
List all configured MCP servers from ~/.claude.json with their configuration. By default, values inside each server's 'env' field are redacted to '***' so this tool can't be used to exfiltrate API keys stored in MCP env. Pass show_env_values=true to reveal them when you legitimately need them.
Inputs: show_env_values.
remove_mcpA
Remove an MCP server from the user configuration (~/.claude/mcp.json).
Inputs: mcp_name (required).
restart_codex_mcp_processA
Restart a Codex CLI MCP server process. Works reliably only for wrapped MCPs (via wrap_codex_mcp_stdio) by killing the wrapped child PID from its pidfile; otherwise you typically need to restart Codex CLI.
Inputs: mcp_name (required).
restart_mcp_processA
Restart the underlying process for an MCP server by finding and killing its process. For wrapped stdio MCPs (via wrap_mcp_stdio), the wrapper will respawn automatically; otherwise you typically need to restart Claude CLI.
Inputs: mcp_name (required).
unwrap_mcp_stdioA
Unwrap a previously wrapped stdio MCP server, restoring its original command. Requires restarting Claude Code to take effect.
Inputs: mcp_name (required).
wrap_codex_mcp_stdioA
Wrap a Codex CLI stdio MCP server (from ~/.codex/config.toml) so it can be restarted reliably. Rewrites the MCP entry to launch mcpgo's Node wrapper that respawns the original command. Restart Codex CLI to take effect.
Inputs: mcp_name (required).
wrap_mcp_stdioA
Wrap a stdio MCP server so it can be restarted reliably. Updates ~/.claude.json to launch a Node wrapper that spawns the original command and respawns it when killed. Requires restarting Claude CLI to take effect.
Inputs: mcp_name (required).
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 11 tools
Each tool has a distinct purpose: listing, retrieving, health-checking, restarting, wrapping/unwrapping, configuring, adding, and removing MCP servers. The two restart tools and the two wrap tools are clearly differentiated by target client (Claude vs Codex), so there's no ambiguity.
All tools follow a consistent verb_noun pattern (list_all, get, check, restart, wrap, unwrap, configure, remove, add). The inclusion of client-specific modifiers (codex, stdio) is consistent and descriptive, and there are no mixed conventions or vague verbs.
With 11 tools, the server is well-scoped for managing MCP server configurations. Each tool covers a distinct operation needed for lifecycle management, and the count is right in the sweet spot for a focused administrative utility.
The surface covers core CRUD operations (add, get, list, remove, configure) plus health and restart. However, there is no unwrap_codex_mcp_stdio tool to complement wrap_codex_mcp_stdio, leaving a minor asymmetric gap that agents would need to work around.
Related MCP Connectors
Augments MCP Server - A comprehensive framework documentation provider for Claude Code
Secure and manage MCP servers through the Speakeasy AI Control Plane from your agent.
Shared memory MCP server for Claude Code, Codex, Gemini CLI and Cursor. Cited, editable.
Create, deploy, and operate MCP servers directly from your GitHub repositories.
Related MCP Servers
- FlicenseNot gradedqualityDmaintenanceEnables dynamic loading, hot-reloading, and orchestration of MCP servers without restarting Claude Code, allowing programmatic tool calling and workflow automation across multiple servers.2-
- AlicenseNot gradedqualityDmaintenanceCreates, manages, and registers custom MCP servers for Claude Desktop. Enables users to generate and configure new MCP servers through natural language.11MIT
- AlicenseAqualityDmaintenanceA meta-MCP server that dynamically manages and invokes any MCP server on the fly, enabling runtime addition, removal, and tool invocation without restarting Claude Code.55 npmMIT
- AlicenseAqualityFmaintenanceAn MCP server for coordinating multiple Claude Code sessions across related projects.11MIT