localhost-mcp
localhost-mcp
Servidor MCP que inspecciona, gestiona y finaliza servidores de desarrollo locales. Deja de adivinar qué hay en :3000.

Se combina con terminal-history-mcp: juntos le dan a tu agente de IA memoria completa de tu entorno de desarrollo: qué ejecutaste y qué se está ejecutando.
Por qué
Todo desarrollador se enfrenta a esto a diario:
Error: listen EADDRINUSE :::3000— ¿qué está ocupando el puerto?5 PIDs de
node/vite/nextolvidados de la semana pasada consumiendo RAMCambiar de proyecto → ni idea de qué servidores de desarrollo siguen ejecutándose
lsof -i :3000,kill -9 <pid>, repetir
localhost-mcp lo convierte en una llamada de herramienta.
Related MCP server: Git Insight MCP
Instalación
npm install -g localhost-mcpConéctalo a Claude Code:
claude mcp add --scope user localhost -- localhost-mcpO a cualquier cliente compatible con MCP. El comando se ejecuta como un servidor MCP stdio.
Herramientas
Herramienta | Propósito |
| Todos los servidores de desarrollo en escucha con puerto, pid, framework, proyecto, tiempo de actividad, memoria, cpu |
| Inspeccionar un solo puerto: quién lo ocupa |
| Finalizar por pid o puerto. Ejecución en seco por defecto; pasa |
| Detectar servidores de desarrollo de larga duración, inactivos o con alto consumo de memoria |
| Por qué el puerto X está ocupado + 5 alternativas libres cercanas |
Salida de muestra
{
"port": 3000,
"pid": 48211,
"process": "node",
"cmdline": "next dev",
"cwd": "/Users/me/code/myapp",
"project_name": "myapp",
"framework": "next.js",
"uptime_seconds": 14523,
"memory_mb": 412,
"cpu_pct": 0.3,
"user": "me"
}Seguridad
kill_serverutiliza ejecución en seco por defecto. Se debe pasarconfirm=true.Se niega a finalizar PIDs < 1000 (procesos del sistema).
Se niega a finalizar procesos fuera de la lista blanca de desarrollo (node, python, ruby, go, deno, bun, php, java, rails, vite, next, etc.).
SIGTERM primero, escala a SIGKILL después de un tiempo de espera de 5s.
Frameworks detectados
next.js, vite, nuxt, remix, astro, webpack-dev-server, esbuild, create-react-app, express, fastify, koa, hono, rails, django, flask, fastapi, uvicorn, gunicorn, deno, bun, php-builtin, jekyll, hugo.
Recurre a la inspección de package.json cuando la línea de comandos es genérica (node server.js).
Soporte de plataforma
macOS — soporte completo (usa
lsof)Linux — soporte completo (usa
lsof+/proc)Windows — solo escaneo básico de puertos (usa
netstat); detección de cwd / framework limitada
Uso de CLI
localhost-mcp list # JSON list of all dev servers
localhost-mcp zombies # JSON list of zombie candidates
localhost-mcp # Start MCP stdio serverConstruir desde el código fuente
git clone https://github.com/hasanjahidul/localhost-mcp.git
cd localhost-mcp
npm install
npm run build
node dist/cli.js listLicencia
MIT
Available Tools
5 toolsfind_zombiesARead-onlyIdempotent
Read-only. Flags dev servers that look abandoned — all three of: uptime > 6h AND CPU < 1% AND memory > 100MB. By default it excludes known always-on noise (VS Code server, language servers, other MCP servers, postgres/redis, sidekiq, etc.); set include_excluded: true to also list those. Each candidate comes with the reasons it matched. This tool never kills anything — pass results to kill_server. Returns { count, candidates[] }.
Inputs: include_excluded. Returns: count, candidates
kill_serverADestructive
Terminates a local dev server by pid or port. DESTRUCTIVE. Safe by default: with no confirm, this is a DRY RUN — it reports what it would kill and changes nothing. Pass confirm: true to actually terminate: sends SIGTERM, waits up to 5s, then escalates to SIGKILL (or SIGKILL immediately if force: true). Refuses PIDs below 1000 and processes that don't look like dev servers. Provide exactly one of pid or port. Returns what was (or would be) killed and the signal used.
Inputs: pid, port, force, confirm. Returns: pid, port, killed, signal, +3 more
list_dev_serversARead-onlyIdempotent
Read-only. Lists every local development server (next, vite, nuxt, remix, astro, rails, django, flask, express, deno, bun, etc.) currently LISTENING on a TCP port. For each: port, pid, process name, command line, working directory, project name, detected framework, uptime, memory (MB), CPU %, and owning user. Uses lsof on macOS/Linux and netstat on Windows; on Windows the cwd/framework fields are limited. Takes no arguments. Returns { count, servers[] }.
Inputs: no named parameters. Returns: count, servers
port_conflictARead-onlyIdempotent
Read-only. Diagnoses an EADDRINUSE situation: given a port, returns the dev server currently blocking it plus 5 free alternative ports nearby. Use it when a listen EADDRINUSE error fires and you want both the culprit and a port to switch to.
Inputs: port (required). Returns: port, blocked_by, alternatives
port_infoARead-onlyIdempotent
Read-only. Inspects a single TCP port and reports the dev server holding it (same fields as list_dev_servers). If nothing is listening, or the listener is not a recognised dev process, returns { port, status: "free" }. Useful for answering "what's on :3000?" before starting or killing something.
Inputs: port (required). Returns: Either a dev-server record (see list_dev_servers) or `{ port, status: "free" }`.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 5 tools
Each tool has a clearly distinct purpose: listing servers, inspecting a single port, finding zombies, killing servers, and diagnosing port conflicts. There is no functional overlap.
Tool names use snake_case, and most follow a verb_noun pattern (find_zombies, kill_server, list_dev_servers). However, port_conflict and port_info are noun-first, which is a minor inconsistency.
With 5 tools, the set is well-scoped for managing local dev servers: query, inspect, cleanup, and conflict resolution. No tool feels superfluous or missing.
The tools cover listing, inspecting, zombie identification, killing, and conflict diagnosis. There is a minor gap in that bulk operations (e.g., kill all zombies) are not directly supported, but the intended workflow is clear.
Maintenance
Related MCP Connectors
2,000+ MCP servers read at source level. Know what one does before you connect. Free, no key.
One MCP server for all your MCP servers: code mode, per-tool permissions, approvals and audit.
Scan any MCP server for tool-poisoning, security, auth & license. Trust score before install.
Self-hosted MCP server: 26 deterministic dev, security, and EVM tools.
Related MCP Servers
- AlicenseAqualityCmaintenanceDevelopers: Search your zsh, bash, or fish shell history from Claude Code, Cline, Cursor, Zed, or any MCP client using tools like search_history (full-text with timestamp/CWD/exit code), recent_in_dir, failed_commands, and command_chains for multi-step sequences. Reindex after new activity. Local-only SQLite FTS5 with secrets redacted before storage.7149 npm5MIT
- AlicenseAqualityDmaintenanceSemantic git queries via MCP. Beyond git log — answer who/what/why about any line, file, or branch with blame, co-change, PR linkage.613 npm2MIT
- AlicenseNot gradedqualityDmaintenanceSee and control the local dev servers your coding agents leave running. Lists listeners with provenance — which agent, terminal and git worktree started each — kills strays, and allocates collision-free ports so parallel agents stop fighting over :3000.MIT
- AlicenseNot gradedqualityCmaintenanceMCP server for managing local dev ports on macOS. It enables AI agents to inspect listening ports, identify owning processes and parent chains, kill processes safely, wait for ports, and report LAN exposure.8 npm1MIT