Mesh LLM browser extension

⚠️ Experimental, best effort, use at your own risk. This is an unofficial, early-stage project with no support and no guarantees. It installs and runs a preview build of mesh-llm on your Mac. Public mode joins an open mesh of other people's computers, so anything you send in chat goes to whichever peer serves it. Don't use it for anything sensitive. It may break, change, or disappear at any time.

Quick start (macOS, Apple Silicon, Chrome or another Chromium browser)

1. Run this in Terminal:

curl -fsSL https://raw.githubusercontent.com/Mesh-LLM/browser-extension/main/install.sh | bash

2. Load the extension. The installer opens chrome://extensions and copies the folder path to your clipboard.

• Turn on Developer mode (top right), then click Load unpacked.
• Press ⌘⇧G, paste with ⌘V, press Enter, and click Open.

3. Click the jellyfish in the toolbar, then click Start Mesh.

To update, click Update in the popup. To remove everything:

curl -fsSL https://raw.githubusercontent.com/Mesh-LLM/browser-extension/main/install.sh | bash -s -- --uninstall

---

What it does

Start, join and share a mesh-llm node from your browser, and chat with the mesh beside any page.

Popup (the control panel)

• Status line: public or private mesh, peer count, and what this Mac is sharing. The toolbar badge shows the peer count: green for public, violet for private. Full details are in the Console.
• Chat opens the Mesh chat (:3131/chat). There's also a small Console link.
• Public / Private
  • Public joins the open mesh (--auto).
  • Private is your own mesh. Copy an invite puts a 24-hour invite on the clipboard. Join with an invite pastes one, then restarts and joins.
• Share compute (on or off), with an optional model. Turning it off makes this Mac a client only.
• Start / Stop / Restart to apply. Settings changes are applied with a restart.
• Start Mesh when the browser opens, Logs, and Update.

Side panel (chat beside any page)

• Streaming chat with history and a choice of model.
• Include this page, Summarise this page, Key points, and Explain my selection.
• Right-click menu: Ask the mesh about "…" and Summarise this page with the mesh.

Install details

What the installer does:

• Everything goes into ~/.mesh-llm-browser, with no sudo.
• It downloads this repo's extension and launcher, plus the latest mesh-llm preview. The mesh-llm download is checked against its published checksum.
• It registers the launcher with Chrome, Brave, Edge, Arc, Chromium and Vivaldi, whichever are installed.
• Read install.sh first if you like. It's short.

Updating: click Update in the popup. It updates mesh-llm, the extension files and the launcher, then reloads the extension. Re-running the install command does the same.

Why Developer mode? On macOS, Chrome only installs extensions from the Chrome Web Store, or unpacked in Developer mode. This project isn't in the store, so it's loaded unpacked. It's the same code you can read here.

If mesh-llm is already running

The extension shows it and works with it: status, Chat, side panel and invites. It never starts a second copy, because the launcher refuses when port 3131 or 9337 is already in use. Stop asks for confirmation before stopping a node it didn't start.

How settings map to mesh-llm

Settings	Command
Public, sharing	serve --auto --model <auto pick or your choice>
Public, not sharing	client --auto
Private, my own mesh	serve --owner-required --trust-policy require-owned --min-node-version 0.76.0 --model <auto>
Private, joined with an invite	serve --owner-required --trust-policy require-owned --join-file ~/.mesh-llm-browser/invite.txt
Private, not sharing	add --client to either private line

• Automatic model, picked by memory tier. Without one, serve --auto just sits on standby:
  • 128 GB or more of memory (with at least 22 GB free): Qwen3.8-27B.
  • 24 GB or more (12 GB free): Gemma 4 12B.
  • 16 GB or more (6 GB free): Qwen3.5-4B.
  • Less than that: no model, so it joins on standby (or as a client in Private mode).
  • A model you type in the popup wins over the automatic pick, and [[models]] in your ~/.mesh-llm/config.toml wins over both.
• Uses your normal ~/.mesh-llm profile, so the identity, config and models are the same as when you run mesh-llm from Terminal. Model files live in the Hugging Face cache, not in ~/.mesh-llm.
• A private mesh needs an owner identity. If the machine has none, the launcher creates one with mesh-llm auth init --keychain. It never replaces an existing one.
• The invite is written to a file only you can read (0600) and passed with --join-file, so it doesn't appear in process listings.

How it works

popup / side panel ──fetch──▶ 127.0.0.1:3131  (/api/status, console, /chat)
                   ──fetch──▶ 127.0.0.1:9337  (/v1/models, /v1/chat/completions)
background.js ──native messaging──▶ ~/.mesh-llm-browser/mesh-llm-host.pl ──fork+setsid──▶ mesh-llm

The launcher (native-host/mesh-llm-host.pl) is plain Perl, which macOS always includes. It accepts only these commands: status, start, stop, restart, logs, update and args (a dry run). Settings arrive as checked fields (connection, share, model, invite) and are turned into a command line from a fixed list. mesh-llm is fully detached, so it keeps running after the browser closes. Logs go to ~/.mesh-llm-browser/mesh-llm.log.

Uninstall:

curl -fsSL https://raw.githubusercontent.com/Mesh-LLM/browser-extension/main/install.sh | bash -s -- --uninstall

This leaves your ~/.mesh-llm profile and downloaded models alone.

TODO

• Firefox: a separate manifest using background.scripts and sidebar_action, signed as a self-distributed .xpi.
• Linux support in the installer.