Meet WebBrain: An Open-Supply, Native-First AI Browser Agent That Reads Pages and Automates Duties in Chrome and Firefox

WebBrain is a free, open-source browser agent for Chrome and Firefox. It reads pages, extracts knowledge, and automates multi-step duties. Not like most browser AI plugins, it might probably additionally run fully on a neighborhood mannequin.

It’s constructed by Emre Sokullu and licensed below MIT. The total supply lives on GitHub. 

Run the agent in opposition to a neighborhood mannequin, and no web page knowledge leaves your machine. Join a cloud API if you need extra functionality.

What’s WebBrain?

WebBrain lives in your browser’s facet panel. In Chrome it makes use of Manifest V3 and the sidePanel API. In Firefox it makes use of Manifest V2 and sidebar_action. Every tab retains its personal dialog historical past.

The extension operates inside your present authenticated session. It sees your logged-in accounts precisely as you do. It shops no knowledge externally and provides no telemetry or accounts.

The plugin ships in English, Español, Français, Türkçe, and 中文. It auto-detects your browser language on first launch.

Ask Mode, Act Mode, and How Actions Truly Hearth

WebBrain has two modes: Ask mode is read-only and can’t change the web page. Act mode can click on, sort, scroll, navigate, and run workflows.

Ask mode reads pages by unusual content material scripts. Act mode is totally different. It drives the web page by the Chrome DevTools Protocol by way of the chrome.debugger API. That produces trusted enter occasions that fashionable websites really honor. It additionally reaches cross-origin iframes and shadow DOM that content material scripts can’t see.

That energy is scoped intentionally. WebBrain attaches the debugger solely when an motion wants it, per tab. Chrome surfaces its normal ‘WebBrain began debugging this browser’ banner whereas connected. Firefox has no CDP equal, so its Act mode is meaningfully weaker.

Temperatures are mounted for predictability. Act mode makes use of temperature 0.15. Ask mode makes use of 0.3. Devoted imaginative and prescient screenshot descriptions use 0.

The Safety Mannequin

Browser brokers run on an adversarial floor. Internet pages can conceal immediate injections that hijack an agent’s conduct. WebBrain’s design addresses this instantly.

The agent begins in read-only Ask mode. It asks earlier than consequential actions. You may disable these prompts within the Permissions settings. They’re on by default.

There may be additionally a UI-first rule for mutations. For something that creates, sends, submits, or buys, WebBrain makes use of the seen UI. It refuses to name REST or GraphQL endpoints instantly for mutations. A per-conversation /allow-api override exists when the UI genuinely fails.

Studying is handled individually. Fetching a README or evaluating costs makes use of background HTTP by the fetch_url and research_url instruments. Studying modifications nothing remotely, so the strict guidelines don’t apply.

Use Circumstances, With Concrete Examples

• Knowledge extraction is the apparent one: Open a catalog and ask: ‘Extract all product names and costs from this web page.’ The agent reads the construction and returns rows. It additionally works with PDFs.
• Analysis summaries are one other: Ask ‘Summarize this text,’ then observe up with a particular query. WebBrain detects paywalls actually and doesn’t attempt to bypass them. It additionally dismisses frequent cookie-consent banners earlier than studying.
• Type filling fits repetitive signups: An non-compulsory Profile auto-fill shops a brief bio in native plaintext. That textual content is shipped to your configured LLM to finish low-stakes types. Maintain essential passwords out of it.
• Automation spans a number of steps: Strive ‘Navigate to github.com and discover trending repositories.’ In Act mode, the agent chains navigation, reads, and clicks.

Maintaining Token Prices Down

Cloud tokens add up on lengthy classes. WebBrain bounds the price in 3 ways.

• Screenshots are resized and iteratively JPEG-compressed earlier than they depart your machine. That retains picture tokens small. 
• Dialog historical past and gear outputs are trimmed oldest-first because the context window fills. 
• You may as well pair an affordable textual content mannequin for planning with a separate imaginative and prescient mannequin for screenshots.

How It Compares

WebBrain sits between browser AI plugins and full agent frameworks. Right here is the plugin comparability, drawn from the venture’s personal documentation.

Frameworks like OpenClaw or Browser-Use are a distinct class. These are developer SDKs for headless pipelines. WebBrain is an end-user extension you drive from a chat panel. You need to use each.

Operating It: Suppliers and Setup

WebBrain helps native and cloud fashions by one interface. Native choices embrace llama.cpp, Ollama, LM Studio, Jan, vLLM, and SGLang. Cloud choices embrace OpenAI, Anthropic Claude, Gemini, Mistral, DeepSeek, and xAI Grok. It additionally helps Groq, MiniMax, Alibaba Cloud (Qwen), Nvidia NIM, and OpenRouter.

A built-in managed possibility, WebBrain Cloud, wants no native setup. It prices $5 per thirty days per system profile below a fair-use coverage. For native use, llama.cpp wants no API key.

Beginning a neighborhood server takes one command:

# llama.cpp — load not less than a 16k-token context window
llama-server -m your-model.gguf -c 16384 --port 8080

# Ollama (OpenAI-compatible) — set the extension-origin env var
OLLAMA_ORIGINS="*" ollama serve
# then set the bottom URL to http://localhost:11434/v1 in settings

Level WebBrain on the endpoint in settings. For a cross-machine vLLM server, allow CORS with –allowed-origins ‘[“*”]’.

The beneficial mannequin is Qwen 3.6 35B (Qwen3.6-35B-A3B). It beat Gemma 4 on the venture’s screenshot benchmark. An RTX 5090 is right; an RTX 4090 works with INT4 AutoRound quantization.

Every supplier is a category that extends BaseLLMProvider. It normalizes to at least one response form:

null 

Key Takeaways

• WebBrain is a free, MIT-licensed AI browser agent for Chrome and Firefox, constructed by Emre Sokullu.
• It runs on native fashions (llama.cpp, Ollama; Qwen 3.6 35B beneficial) or any cloud API — no web page knowledge leaves your machine when native.
• Ask mode reads pages read-only; Act mode clicks and kinds by way of the Chrome DevTools Protocol for trusted enter occasions.
• Safety-first by design: begins read-only, approves consequential actions, and makes use of the UI as an alternative of direct API requires mutations.
• Free ceaselessly self-hosted, or $5/month per system profile for the managed WebBrain Cloud below truthful use.

Interactive Explainer with Demo

Demo-1