Mcp Launcher For Claude Code

MCP Launcher for Claude Code

Per-session MCP server loadouts — launch lean, measure the cost.

A compact Windows desktop app (PySide6) that keeps Claude Code's context light by launching each session with only the MCP servers it needs.

> Unofficial, community-made tool. Not affiliated with or endorsed by Anthropic.

The real cost of MCP servers isn't CPU/RAM (an idle server costs almost nothing) — it's context: every connected server injects all its tool definitions into the model's context on every request. This app lets you pick a loadout, see its token weight, and launch a Claude session that loads exactly those tools.

> A configured loadout — 2 of 7 servers selected, each server's tool count and > token weight shown, with the running total in the header. Tick the servers a > task needs, then Launch a Claude session scoped to exactly those.

On first run with no servers configured yet, it starts clean and seeds from your Claude Code config:

Download

Grab the latest standalone build from the [Releases](../../releases) page — unzip it, open the MCP Launcher folder, and run MCP Launcher.exe. It's self-contained: no Python or other install needed. Prefer to run from source instead? See [Run from source](#run-from-source) below.

Requirements

• Windows 10/11 (64-bit).
• The Claude Code CLI (claude) installed and on PATH — this app

launches it, it doesn't replace it. Only needed when you actually Launch a session; the app opens fine without it.

• Python 3.12+ (only when running from source; the packaged build bundles it).

Run from source

Just double-click run.bat. On the first run it creates the venv, installs the dependencies from requirements.txt, and then launches; every run after that skips straight to launching. (It needs Python 3.12+ on PATH to do the one-time setup; it'll tell you if it can't find one.)

Prefer to set it up by hand? That's all run.bat does:

python -m venv venv
venv\Scripts\python -m pip install -r requirements.txt
run.bat

The app runs windowless via pythonw (no console). If you launch run.pyw directly with a Python that doesn't have the dependencies installed, it shows a message box pointing you back to run.bat instead of failing silently.

How it works

• Catalog — keeps your MCP servers in

%LOCALAPPDATA%\mcp-launcher\store.json, seeded on first run from ~/.claude.json (both user-scope and per-project servers).

• Loadout — tick the servers you want for a task. The header shows the

total estimated token cost of that selection.

• Launch — opens a new terminal running

claude --mcp-config .json --strict-mcp-config in a folder you choose. --strict-mcp-config makes that session use only those servers, ignoring the global ~/.claude.json entirely.

Because each session is launched with its own private config, two sessions never share a mutable MCP config, so they can't clobber each other (and a running session can't overwrite your changes). Verified: a strict session with a one-server preset sees only that server; with an empty preset it has no MCP tools at all.

• Presets — save a selection as a named loadout (e.g. 3D, Audio,

Unity, Minimal) and recall it from the dropdown.

• Measure context cost — spins each server up briefly, lists its tools, and

estimates the tokens its definitions add per request. Per-server badges are colour-graded (green/amber/red) and the header sums the current loadout.

• Set as global default — writes the current selection into ~/.claude.json

so a plain claude launch (without this app) loads exactly that set. Atomic write, timestamped backup in %LOCALAPPDATA%\mcp-launcher\backups (last 15 kept), and all other config keys preserved.

• Add / Edit / Delete servers — edit shows the raw JSON record (with an

"OVERWRITE?" confirm); delete confirms too.

What it touches

• Reads, and (only via Set as global default) writes, the mcpServers blocks

of Claude Code's config (~/.claude.json, or $CLAUDE_CONFIG_DIR/.claude.json if you've relocated it). It never reads or modifies any other key (prompts, history, auth, project settings stay untouched).

• Its own data (catalog, presets, backups, generated launch configs) lives under

%LOCALAPPDATA%\mcp-launcher.

Security & privacy

This is a local utility with a deliberately small footprint:

• No network, no telemetry, no listeners. It opens no sockets, makes no

outbound requests, and phones nothing home. Everything happens on your machine.

• It reads only the mcpServers blocks of ~/.claude.json — never your

prompt templates, conversation/project history, OAuth tokens, or other keys. The safety backup it writes before Set as global default contains only those MCP blocks, so no personal data is ever copied into the app folder.

• It runs MCP server commands — that's inherent to MCP. Claude launches them

per session, and Measure context cost launches each briefly to read its tools (it asks first). Only add servers you trust. When you add or edit a record, the dialog validates its shape and shows the exact command + args, and you confirm before it's saved — nothing runs unseen.

• Launching is injection-safe by construction. Claude and each argument are

handed to the OS as a list (no shell command string is built), the working folder is passed via the process cwd parameter and verified to exist, and the config path is an app-generated file with a sanitized name — so a crafted folder name or path can't smuggle in an extra command.

• Secrets stay local. A server's env may hold API keys; those live in

store.json and the generated launch files under %LOCALAPPDATA%\mcp-launcher (a per-user, OS-protected location) — the same exposure as your existing ~/.claude.json. Don't commit or share those files; .gitignore already keeps the repo clean of local data and the bundled venv.

No personal or machine-specific data is embedded in the source or this README.

Platform note

Windows-only for now (uses Windows Terminal / PowerShell and %LOCALAPPDATA%). The launcher could be extended for macOS/Linux later.

Project layout

mcp_launcher/
  paths.py          app data locations + bundled-asset/AppUserModelID ids
  claude_config.py  safe atomic read/write of ~/.claude.json + launch-config writer
  store.py          catalog of servers + presets + selection (store.json)
  prober.py         measure each server's tool/token context cost (FastMCP)
  launcher.py       build a --mcp-config file and open a Claude session
  app.py            entry point + global excepthook + taskbar identity
  assets/
    icon.ico        multi-resolution "M" app icon (window/taskbar/.exe)
  ui/
    branding.py     the "M" logo: live QIcon + icon.ico generator
    style.py        dark theme (QSS)
    record_row.py   one server row (checkbox + cost badge + edit/delete)
    dialogs.py      add/edit record, confirmations
    main_window.py  loadout list, presets, launch
run.pyw / run.bat   launchers (run.bat self-bootstraps the venv on first run)
build_exe.bat       build a standalone .exe (PyInstaller, icon wired in)
requirements.txt    runtime dependencies
docs/               screenshots used in this README

License

[MIT](LICENSE) © Maxii-VR.

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Maxii-VR
• Source: Maxii-VR/mcp-launcher-for-claude-code
• License: MIT

Install and usage instructions live in the source repository linked above.