Lmstudio Local Tools

lmstudio-local-tools

GitHub repo License: MIT

A local-AI fallback stack for LM Studio: terminal and browser control via MCP, a stdlib-only terminal chat client, and a keyless web search server. Point any tool-calling-capable local model at your shell, your real Chrome, and the open web — no cloud API, no API keys, nothing leaves your machine.

What's included

qwen	stdlib-only Python REPL that talks to LM Studio's local server API, including MCP tool calls. Works with any tool-calling-capable model.
rg_websearch_mcp.py	keyless MCP server exposing web_search(query, max_results=5), backed by Bing's HTML results page — no API key required.
mcp.json.example	template MCP server config for ~/.lmstudio/mcp.json.
scripts/ + launchagents/	optional LaunchAgent to keep the LM Studio local server running in the background.

MCP servers used:

• terminal → @wonderwhy-er/desktop-commander — shell command execution, file read/write/edit/search.
• browser → @playwright/mcp --extension — controls your real running Chrome via a bridge extension.

Both run via npx, no persistent install required.

Setup

1. Copy the MCP config. Copy mcp.json.example to ~/.lmstudio/mcp.json (merge with any existing entries), filling in the placeholders:

   • PLAYWRIGHT_MCP_EXTENSION_TOKEN — shown when you install the Playwright Extension in Chrome.
   • The absolute path to rg_websearch_mcp.py for the websearch entry.

2. Reload MCP servers in LM Studio. Open LM Studio → left sidebar → Program (plug/MCP icon). You should see terminal, browser, and websearch listed; hit refresh or restart LM Studio if not.

3. Load a tool-calling capable model from LM Studio's model picker, and enable the terminal/browser/websearch MCP tools for the chat.

4. Enable server authentication (Developer tab → Local Server → Server Settings): turn on Require Authentication, then Manage Tokens → Create Token, and export it:

   export LM_API_TOKEN="<your token>"
   

   This must be on before "Allow calling servers from mcp.json" will take effect — enabling only the mcp.json toggle without authentication first silently does nothing.

Usage

cd lmstudio-local-tools
export LM_API_TOKEN="<your token>"
./qwen                          # interactive REPL, both tools on
./qwen --no-browser              # terminal tool only
./qwen --no-terminal              # browser tool only
./qwen --model "<model-id>"
./qwen --show-reasoning           # show the model's reasoning blocks
./qwen -p "list files in ~/Desktop"   # one-shot, non-interactive

$ ./qwen
model: qwen/qwen3.8-27b
tools: ['mcp/terminal', 'mcp/browser']
type 'exit' or Ctrl-D to quit

> list the files on my Desktop
[tool] terminal.list_directory({"path": "~/Desktop"})
  -> Portfolio, Screenshots, notes.txt, ...
Your Desktop has a Portfolio folder, a Screenshots folder, and a notes.txt file.

The client keeps the conversation stateful across turns, prints each tool call it makes with a truncated output preview, then the model's final reply. Type exit or Ctrl-D to quit.

Want to just type qwen from anywhere?

ln -s "$(pwd)/qwen" ~/.local/bin/qwen   # make sure ~/.local/bin is on your PATH

Then qwen works as a plain command in any terminal, no cd or python3 needed.

Test the web search server standalone:

python3 rg_websearch_mcp.py --self-test "your query here"

Isolated browser profile (optional)

If you'd rather not install the Playwright extension, edit the browser entry in mcp.json to use a separate persistent Chrome profile instead of your real one:

"args": ["-y", "@playwright/mcp@latest", "--browser", "chrome", "--user-data-dir", "~/.lmstudio-browser-profile"]

Auto-start the LM Studio server

Copy scripts/ensure-lmstudio-server.sh to ~/.local/bin/ and launchagents/com.lmstudio.server-autostart.plist to ~/Library/LaunchAgents/, updating the paths inside each to match your username, then:

launchctl load ~/Library/LaunchAgents/com.lmstudio.server-autostart.plist

This runs lms server start at login and every 5 minutes thereafter.

Safety notes

• terminal gives the model real shell access — full read/write/delete on your filesystem and the ability to run any command. Leave LM Studio's per-tool-call confirmation dialog on (neverAskForToolConfirmation: false in ~/.lmstudio/settings.json), and don't approve a command you don't understand.
• browser with --extension can act in your real, logged-in browser. Don't approve actions on sensitive tabs you haven't reviewed.
• qwen has no per-tool-call confirmation of its own — the API executes tool calls immediately once the server setting above is on.
• Local models are more likely than a frontier hosted model to misread instructions or run a destructive command by mistake. Review before approving anything destructive.

Notes on the web search backend

rg_websearch_mcp.py scrapes Bing's plain HTML results page rather than DuckDuckGo, because DuckDuckGo's html.duckduckgo.com / lite.duckduckgo.com endpoints return an anti-bot CAPTCHA for a bare scripted request. Scraping a search engine's HTML is inherently a "verify it still works" integration — if Bing starts blocking requests too, check the page's current markup in a browser and adjust the regexes accordingly.

License

MIT © 2026 RouterGlock — see LICENSE.