TheRadar

Still · Open source · MIT

MCP servers that sit still until you call them.

Still is a small command-line tool. It turns each of your MCP servers into a short skill file that Claude reads, and it starts the real server only when a tool is called. The rest of the time, nothing runs.

The problem

Every MCP server you add is a process that never leaves.

You install one MCP server. Then five. Then twelve. Each is a long-running process. Open one Claude Code session, you get a copy of every server. Open three sessions, you get three copies. They don't share. Most of them sit idle, waiting for a call that never comes.

1 session

0

background processes

3 sessions

0

background processes

5 sessions

0

background processes

The arithmetic is linear. A typical power-user setup with about a dozen MCP servers and a few language servers lands near these numbers. Most of those processes will never service a request before the session closes.

How it works

Claude reads a note. The server starts only when it is needed.

Still installs one small CLI — mcp — and writes a thin skill file for each server you've registered. The skill is docs Claude reads to know what tools exist and how to call them. When Claude calls one, mcp spawns the server, runs the call, and the server exits.

always loaded

Skill file

Docs Claude reads

on call

mcp CLI

Spawns the server

ephemeral

MCP server

Runs the call, exits

Three commands:

1 · pull in what you already have
mcp import
2 · generate one skill per server
mcp distill --all
3 · call a tool: server spawns, runs, exits
mcp airtable list-records --base-id appXXXXXXXXXX --table "Tasks"

Install

Two lines to install.

drop the skill into ~/.claude/skills
git clone https://github.com/the-radar/still.git ~/.claude/skills/still
put the CLI on your path
ln -s ~/.claude/skills/still/mcp ~/.local/bin/mcp

Python 3.9+. No third-party packages required. MIT-licensed. The README in the repo covers SSE servers, hardening notes, and the registry format.