Chat with any page. Nothing leaves your machine.

An Ollama browser extension connects the web page you're reading to a large language model running on your own computer. GlassBar is one: a frosted-glass chat bar docked to the bottom of every tab that sends the page's title, URL, and visible text — plus your question — to Ollama or LM Studio on localhost, and streams the answer back in place. There is no cloud dependency, no account, and no API key. With a local model, nothing you read or ask ever leaves your machine.

This page covers the full setup, including the one step most people miss — the OLLAMA_ORIGINS environment variable — and an honest look at when a local model is the right tool.

How it works

GlassBar has no backend and no proxy. When you ask a question, the extension gathers the page's title, URL, and visible text and sends the request directly from your browser to whichever provider you've configured. Select Ollama or LM Studio and that provider is a server on your own machine — the request goes to localhost and never touches the internet. There is no middleman to log, sell, or leak what you read, and the developer receives no user data. The details are in the privacy policy.

The bar itself stays out of the way: it rests as a small pill at the bottom of the page, expands on hover, pins with a click, and tucks away with Esc. Alt+J toggles it from the keyboard. Answers stream in as real formatting — headings, lists, code — not raw markdown. It works on Firefox (desktop and Android, Firefox 142+) and runs on Chrome, Edge, and Brave from the same codebase; the Chrome Web Store listing is coming soon.

Set up Ollama with GlassBar

1. Install Ollama

Download Ollama from ollama.com for macOS, Windows, or Linux. Once it's running, it serves a local API at http://localhost:11434.

2. Pull a model

Grab a model that fits your RAM — llama3 is a reasonable default:

ollama pull llama3
ollama run llama3   # quick sanity check, then /bye

3. Allow browser-extension origins

This is the step that trips people up. Extensions call the Ollama API from their own origin — moz-extension://… in Firefox, chrome-extension://… in Chromium — and Ollama only accepts browser requests from localhost origins by default, rejecting everything else with 403 Forbidden. As of July 2026 the fix is still the OLLAMA_ORIGINS environment variable, documented in the Ollama FAQ. Set it before Ollama starts:

# macOS (menu-bar app)
launchctl setenv OLLAMA_ORIGINS "moz-extension://*,chrome-extension://*"
# then quit and reopen Ollama

# Linux (systemd)
sudo systemctl edit ollama.service
#   add under [Service]:
#   Environment="OLLAMA_ORIGINS=moz-extension://*,chrome-extension://*"
sudo systemctl daemon-reload && sudo systemctl restart ollama

On Windows, quit Ollama, add OLLAMA_ORIGINS with the same value as a user environment variable in Settings, then relaunch. The wildcard covers extension IDs, which differ per install; it does not open the server to arbitrary websites.

4. Point GlassBar at localhost

Install GlassBar from addons.mozilla.org, open its settings, and select Ollama as the provider. Pick your model, open any page, and ask. No key, no sign-up — the free tier includes local models, every provider, and page-aware chat, permanently.

Using LM Studio instead

LM Studio works the same way. Load a model, start the local server from its developer view — it serves an OpenAI-compatible API on http://localhost:1234 by default — and enable CORS in the server settings so the extension can reach it. Then select LM Studio in GlassBar's settings. Same bar, same streaming answers, same zero-cloud path. GlassBar also accepts any OpenAI-compatible endpoint, so other local servers work through that option too.

GlassBar vs Page Assist vs Ollama Client

Two well-known Firefox extensions already talk to Ollama. As of July 2026: Page Assist is an open-source web UI for local models that opens as a sidebar or a full tab, and supports Ollama plus OpenAI-compatible providers. Ollama Client is local-first by design — Ollama, LM Studio, and llama.cpp-compatible servers — with chat in a side panel or popup. Both are solid tools.

GlassBar's difference is where the chat lives and how far it reaches. The bar sits in the page itself, docked to the bottom of every tab, so the answer streams right over what you're reading — no sidebar, no separate tab. And the same bar takes cloud keys: Claude, GPT, Gemini, OpenRouter, Grok, Groq, DeepSeek. When a question outgrows your local model, you switch providers instead of switching tools.

Competitor details from their AMO listings and documentation, as of July 2026. GlassBar is not affiliated with either project.

When local beats cloud — and when it doesn't

Local wins on three counts. Privacy: contracts, medical pages, internal docs — with Ollama or LM Studio, the page text goes to your own hardware and nowhere else. Offline: once a page is loaded, the chat works with no connection at all. Cost: no per-token bill; a pulled model answers as many questions as you like on hardware you already own.

The honest limit is model quality. An 8B model on a laptop is not Claude or GPT — long, dense pages and subtle reasoning still favor the big cloud models. GlassBar treats that as a per-task choice rather than a commitment: each provider keeps its own key and model selection, so you can read the sensitive page on llama3 and send the hard question to Claude. For the cloud side of that setup, see ChatGPT sidebar for Firefox; for everything else, the FAQ.

All of the above is free. The one paid feature is Pro — $15 one-time, never a subscription — which adds Saved Chats: conversations persist per page and auto-restore when you revisit, history is searchable, and any chat exports to Markdown or JSON.

FAQ