Claude Fedora Bridge

There's no Claude desktop app for Linux. So if you're on Fedora and you use Claude in the browser, Claude can't actually touch your machine — it can't read a file, run a build, check why a service is failing, or take a screenshot to see what you're looking at.

This is my fix for that. It's a small local server that speaks MCP and hands the browser version of Claude a set of tools for driving this box: look at the system, read and write files in a workspace, build and test projects, read logs, launch apps, grab a screenshot, and ask for root when it genuinely needs it (you get a normal Fedora password prompt and decide).

The browser stays the UI. Claude stays the brain. This just gives it hands.

  Claude in your browser  ──MCP over HTTPS──►  this bridge  ──►  your Fedora machine
       (the thinking)                          (the tools)        (stuff actually happens)

What you need

• Fedora (it leans on journalctl, pkexec, lspci, etc. — should work on most systemd distros with small tweaks)
• python3 and ssh (both already there on a stock Fedora install)
• A paid Claude plan (Pro/Max/Team). Custom connectors are a paid feature, and Anthropic's servers call your bridge from the cloud, which is why it has to be reachable over a public URL.

No virtualenv. The deps go straight into your user account with pip --user.

Getting started

git clone https://github.com/Cx188/Bridge-MCP-for-Claude-Users-on-Fedora.git
cd Bridge-MCP-for-Claude-Users-on-Fedora
./run.sh

First run installs two Python packages (mcp, uvicorn) into your account, opens a public tunnel, and prints something like:

● connected
Paste this into claude.ai → Settings → Connectors:
   https://abcd-1234.a.pinggy.link/mcp-XXXXXXXX

Copy that URL, go to claude.ai → Settings → Connectors → Add custom connector, paste it, save. Claude picks up the tools and you're good. Ask it something like "what's using all my disk?" or "build the project in ~/ClaudeWorkspace/foo and tell me why the tests fail."

Keep the terminal window open — that's the on switch. Close it (or Ctrl+C) and everything stops immediately: the server, the tunnel, the public URL. There's no separate disconnect step, and nothing keeps running in the background or starts at boot.

Prefer clicking an icon? Run ./install.sh once and you get a Connect launcher on your desktop.

About that URL

The tunnel is Pinggy over SSH on port 443, because that's the one thing that reliably gets out of locked-down networks. The catch with the free tier: the URL changes every time you start it, and a session caps out around 60 minutes, so you re-paste occasionally. If that annoys you, a reserved domain from Pinggy Pro or ngrok gives you a URL that never changes — same setup, just swap the tunnel command in run.sh.

What Claude can do once it's connected

• Look around — OS/CPU/memory/disk overview, full hardware inventory, a quick health check (failed services, disk/memory pressure, heavy processes), and read the systemd journal with filters.
• Files — list directories, read files, write files, spin up a new project folder with git initialised.
• Build & test — run any command, or run_build / run_tests which sniff out cargo, npm, meson, cmake, make, or pyproject and do the right thing.
• Desktop — take a screenshot (so it can literally see your screen), launch an app, pop a notification.
• Root, with your say-so — when something needs admin rights it calls request_root_action, which triggers a Polkit dialog on your screen. You type your password or you don't. Claude never gets root on its own.

How safe is this, honestly

Let's be real about it, because you're pointing the internet at your laptop:

• Reads go anywhere you can read — except credential stores. ~/.ssh, ~/.gnupg, keyrings, id_rsa, .netrc, recovery codes and friends are blocked outright.
• Writes and deletes are locked to one workspace (~/ClaudeWorkspace by default) unless you deliberately run in full-access mode.
• Root is never automatic in the normal mode. It's always a pkexec/Polkit prompt that you approve.
• The public URL is the lock. The path has a random secret in it (/mcp-<secret>, stored in ~/.config/claude-fedora-bridge/url-secret). Anyone with the full URL can reach the tools, so treat it like a password — don't paste it in public. Rotate it any time with --rotate-secret.
• It only runs while the window's open. When you're done, close it.

If you want to hand Claude the whole machine with no prompts — the "run everything as root" idea — that exists too, but you have to opt in explicitly and start it as root:

sudo PYTHONPATH=src python3 -m claude_fedora_bridge.server --full-access

Don't do that on a machine you care about. Use a VM.

Running it by hand

run.sh is just convenience. If you want to run the server bare (say, behind your own tunnel):

pip install --user -r requirements.txt
PYTHONPATH=src python3 -m claude_fedora_bridge.server

It binds to 127.0.0.1:8747 and prints the local endpoint. Point whatever tunnel you like at that port.

Stuff I might add

Real GUI automation (clicking and typing into windows, not just screenshots), and maybe making the same tools work with other browser AI agents. No promises.

Notes

Not affiliated with Anthropic — just a thing I built to make my own setup less annoying. MIT licensed, do whatever you want with it.