Open-source AI browser agent for Chrome and Firefox. Chat with any web page, automate browser tasks, and run multi-step agent workflows — powered by your choice of LLM.

• Page Reading — Extracts text, links, forms, tables, and interactive elements from any page
• Browser Actions — Click, type, scroll, navigate, and interact with page elements
• Ask / Act Modes — Read-only mode by default, full agent mode with confirmation
• Plan before Act — Act mode can generate a structured plan, show it for approval, then pin the approved plan to the scratchpad before tools run
• Multi-Step Agent — Autonomous task execution with tool-use loops (configurable, default 130 steps)
• Continue from Limit — When the agent hits the step limit, click Continue to keep going
• Multi-Provider LLM — Supports local and cloud models:
  • WebBrain Cloud 1.0 (cloud, default) — Built-in managed cloud option; no local setup required
  • llama.cpp (local) — No API key needed. Also Ollama, LM Studio, Jan, vLLM, and SGLang
  • OpenAI (GPT-5.5, etc.)
  • Anthropic Claude (native API)
  • Google Gemini, Mistral AI, DeepSeek, xAI Grok, Groq
  • MiniMax, Alibaba Cloud (Qwen)
  • Cloudflare Workers AI, Nvidia NIM
  • OpenRouter (default model: stepfun/step-3.7-flash; access 100+ models)
• Onboarding Wizard — First-launch walkthrough covering Act mode safety and provider setup
• Side Panel UI — Clean chat interface that lives alongside your browsing
• Per-Tab Conversations — Each tab has its own chat history
• Streaming — Real-time token streaming from all providers
• Smart Context — Token-aware auto-compaction (summarizes older turns once the conversation nears the model's context window, with a visible "Context automatically compacted" notice), tool result limits, and emergency overflow recovery
• Browser History Control — Act mode can use native go_back / go_forward history tools instead of CSP-sensitive page JavaScript
• API Shortcut Hints — Repeated clicks that fire the same XHR/fetch request can surface a matching fetch_url suggestion while preserving the UI-first and /allow-api mutation policy
• Custom Skills and Skill Tools — Settings → Skills can import trusted skill text or URLs; skills may also expose HTTP or download-job tools through a webbrain-tools manifest. FreeSkillz.xyz is enabled by default for YouTube transcripts and public media downloads, and can be removed.
• Copy Support — Copy buttons on code blocks and full messages
• Page Inspection Banner — Visual indicator when the agent is interacting with the page
• Stop Button — Abort the agent mid-execution at any time
• Deterministic Act Mode — Act mode uses temperature 0.15 for browser-control decisions; Ask mode uses 0.3, and dedicated vision screenshot descriptions use 0

git clone https://github.com/webbrain-one/webbrain.git

1. Open Chrome → chrome://extensions/
2. Enable Developer mode (top right)
3. Click Load unpacked → select the webbrain/src/chrome folder

git clone https://github.com/webbrain-one/webbrain.git

1. Open Firefox → about:debugging#/runtime/this-firefox
2. Click Load Temporary Add-on
3. Navigate to src/firefox/ and select manifest.json

Note: Temporary add-ons are removed when Firefox restarts. For permanent installation, the extension needs to be signed via addons.mozilla.org.

# Using llama.cpp
llama-server -m your-model.gguf --port 8080

# Or using Ollama (OpenAI-compatible)
ollama serve
# Then set base URL to http://localhost:11434/v1 in settings
# Or run: ollama launch webbrain --model <model>

# Or using Jan (OpenAI-compatible)
# Start Jan's local API server and use http://localhost:1337/v1

# Or using vLLM / SGLang (OpenAI-compatible)
vllm serve your-model --port 8000
python -m sglang.launch_server --model-path your-model --port 30000

WebBrain supports Ollama today through the local OpenAI-compatible provider. A new ollama launch webbrain --model <model> handoff can also configure WebBrain automatically, but it is not integrated into upstream Ollama yet. For now, try it from the codex/ollama-webbrain-launch-handoff branch of esokullu/ollama; we hope Ollama will integrate it upstream.

git clone https://github.com/esokullu/ollama.git
cd ollama
git switch codex/ollama-webbrain-launch-handoff
cmake -S . -B build -G Ninja -DOLLAMA_MLX_BACKENDS=
cmake --build build --parallel 8

OLLAMA_ORIGINS="chrome-extension://*,moz-extension://*" ./ollama serve
./ollama launch webbrain --model <model>

Context window: For reliable agent runs, load a local model with at least a 16k-token context window (the usable minimum). 8k can work with Compact mode enabled (Settings → per-provider checkbox); 4k is too small to hold the system prompt + tool schemas. WebBrain auto-compacts the conversation as it nears the window — it assumes 16k for local models unless you set an explicit context size, so give the model server (e.g. llama-server -c 16384) enough room.

Click the WebBrain icon → the side panel opens. Type a message like:

• "Summarize this page"
• "Find all links about pricing"
• "Fill in the search box with 'AI agents' and click Search"
• "Navigate to github.com and find trending repositories"

Click the gear icon or go to the extension's Options page to configure:

Display Settings:

• Verbose Mode — Show full tool call JSON (off by default)
• Auto-screenshot — Provide visual context when DOM/page reads are insufficient
• Max Agent Steps — Configurable step limit (5-200, default 60)
• Plan before Act — Optionally generate and review a structured Act-mode plan before browser tools run (try mode by default; explicit off is preserved)

Skills:

• FreeSkillz.xyz ships enabled by default and exposes read_youtube_transcript, resolve_public_media, and download_public_media through its skill manifest; remove it from Settings → Skills if you do not want those tools available.
• Imported skills are copied into browser local storage and appended to the agent's system prompt when enabled.
• A skill can expose read-only HTTP tools or short-lived download-job tools with a fenced webbrain-tools JSON manifest. Importing a skill is the trust boundary for its declared HTTPS endpoint; download-job skill tools still run in Act mode and use the normal Downloads permission gate before saving files.
• Tool results from third-party content should be marked resultPolicy: "untrusted" so they are wrapped as data, not instructions.

Providers:

src/chrome/                        src/firefox/
├── manifest.json (MV3)            ├── manifest.json (MV2)
├── src/                           ├── src/
│   ├── background.js              │   ├── background.js (+ background.html)
│   ├── agent/                     │   ├── agent/
│   ├── content/                   │   ├── content/
│   ├── providers/                 │   ├── providers/
│   ├── network/                   │   ├── network/
│   ├── trace/                     │   ├── trace/
│   ├── ui/                        │   └── ui/
│   └── offscreen/                 ├── styles/
├── styles/                        ├── icons/
└── icons/                         └── LICENSE

web/
├── index.html
├── privacy.html
└── vercel.json

Key difference: Chrome uses Manifest V3 (service worker, chrome.scripting, sidePanel API), Firefox uses Manifest V2 (background page, browser.tabs.executeScript, sidebar_action).

Deeper docs live in docs/: architecture, site adapters, providers and models, security model, prompt-injection defense, privacy and data flow, accessibility tree and refs, localization, adding a tool, and test scenarios.

Enabled skills can append additional tool schemas at runtime. For example, the bundled FreeSkillz.xyz skill exposes read_youtube_transcript for YouTube transcripts plus resolve_public_media / download_public_media for public media URLs. These skill tools are not hard-coded in the static table above: if the skill is removed or renamed, the tool disappears or appears under the manifest's declared name.

Compact mode is a reduced tool set + shorter system prompt designed for small local models (2B-8B). In both Chrome and Firefox builds, it cuts the Act-mode schema from 40+ tools to about 20, reducing decision surface and hallucination. Enable it per-provider in Settings (checkbox on local providers; off by default).

Shadow DOM note: The accessibility tree only traverses light DOM. On Web Component-heavy pages (Stripe, Salesforce, Shopify), use get_interactive_elements (pierces open shadow roots) or get_shadow_dom / shadow_dom_query for targeted reads.

The fetch_url and research_url tools also ship as a standalone LM Studio plugin at webbrain/web-tools, for users who want web-fetching tool-use inside LM Studio chats without running the full browser extension. Pure Node, no headless browser.

lms clone webbrain/web-tools

Source: lmstudio-plugin/.

WebBrain accepts slash commands as the first thing on a line in the input box. Type /help to see the list inside the panel.

The default UI-first rule exists because API actions are invisible (you don't see what's being sent), often require separate auth tokens you may not have configured, and can have a much larger blast radius than a visible mis-click. Only use /allow-api when you've decided you want that tradeoff for a specific job.

Chrome side panel shortcuts work when the WebBrain side panel has focus.

• Firefox is meaningfully weaker than Chrome. Firefox has no equivalent to Chrome DevTools Protocol via chrome.debugger, so several Chrome-only features are missing in the Firefox build:
  • Click/type goes through the content-script path (document.querySelector + el.click()) instead of CDP Input.dispatchMouseEvent. This means no shadow-DOM piercing, no real trusted mouse events (some React/Vue handlers won't fire), no closed-shadow-root traversal, and no resolveSelector retry budget.
  • No SPA-navigation-aware retry extension.
  • No conversation persistence across background restarts.
  • No CDP screenshots. Auto-screenshot uses tabs.captureVisibleTab instead, which works for active tabs only and at slightly lower quality.
  • No closed shadow root support for read/extract tools.
  • Site adapters, vision detection, loop detection, the auto-screenshot loop, and the opt-in compact prompt/tool set are mirrored to Firefox.
• SPA navigation detection in Firefox. Some single-page applications may not trigger content-script re-injection after client-side navigation.
• Firefox temporary add-on — Firefox requires the extension to be loaded as a temporary add-on during development, which is removed on restart.

See CHANGELOG.md for the full version history. Recent highlights include Plan before Act, native browser-history tools, repeated-click API shortcut hints, WebBrain Cloud 1.0, scheduled tasks, compact-mode improvements, and native PDF reading.

• Conversation export/import — Save and load chat histories (only export added, import not planned)
• Custom skill tools — User-imported skills can expose read-only HTTP and download-job tools via webbrain-tools manifests
• Keyboard shortcuts — Hotkeys for opening panel, sending messages, switching modes
• Context menu integration — Right-click → "Ask WebBrain about this"
• Auto-screenshot vision context — Send captured viewport context to multimodal models for visual understanding
• Chrome Web Store / Firefox AMO — Official store listings

1. Create a new class extending BaseLLMProvider in src/providers/
2. Implement chat() and optionally chatStream()
3. Register it in src/providers/manager.js

All providers normalize to a common response format:

{ content: string, toolCalls: Array|null, usage: Object|null }

MIT — built by Emre Sokullu