An AI coworker that actually uses your browser

What it does

The extension sits adjacent to your tabs and reads the current page through the accessibility tree—text, links, forms, and PDFs—answering questions in Ask mode or physically interacting with the site in Act mode by clicking, typing, navigating, and downloading. A Dev mode adds page source, network logs, and reversible edits for debugging or deeper inspection. You can point it at a local llama.cpp or Ollama instance, a cloud API from a catalog of 103 built-in providers, or use the managed default that requires no setup.

The interesting bit

Instead of scraping HTML with brittle selectors, the agent consumes the browser’s accessibility tree, giving it a structured, semantic view of interactive elements. The tool tier system—Compact, Mid, or Full—adjusts how many capabilities the model sees based on its context window, so a small local 8k model won’t drown in shadow-DOM tools it cannot handle. There is also a planning stage where the agent can draft a structured plan for user approval before it starts clicking.

Key highlights

• Runs fully offline against local OpenAI-compatible servers (llama.cpp, Ollama, vLLM, LM Studio) or via 103 pre-configured cloud providers.
• Three permission modes: Ask (read-only), Act (click/type/navigate), and Dev (source inspection, console, network, reversible edits).
• Multi-step autonomous loops with a default limit of 130 steps, plus scheduled tasks and page watches that poll for conditions.
• Per-tab conversation history and deterministic temperature defaults—as low as 0 for vision descriptions—to keep browser control predictable.
• Workflows can be saved, exported, and re-run without repeating the full reasoning chain.

Caveats

• The Firefox build is explicitly weaker than Chrome: it lacks shadow-DOM piercing, trusted mouse events (so some React/Vue handlers may not fire), CDP screenshots, and SPA-navigation-aware retries. Some single-page apps may fail to re-inject content scripts after client-side navigation.
• Local models need at least a 16k-token context window; 8k only works with the Compact tool tier, and 4k is too small for the system prompt plus tool schemas.

Verdict

Developers who want an open-source browser agent with local-LLM support and fine-grained permission modes should look here. If you live entirely in Firefox or rely heavily on complex SPAs with shadow DOM, the experience will be noticeably degraded.