CLAUDEZILLA by Boot Industries

Claudezilla brings native Firefox browser automation to AI coding harnesses: Claude Code, Pi Agent, Mercury Agent, omp, Hermes Agent, and other MCP clients.

Control Firefox directly from your terminal: navigate pages, click elements, fill forms, capture screenshots, and analyze page content. A Google-free, telemetry-free alternative to Chrome-based browser automation.

FEATURES:
- Multi-harness support: automated registration for Claude Code and omp; the installer prints ready-to-paste configuration for Hermes, Pi and Mercury
- Full tool catalog by default, with tool titles and read-only/destructive annotations so harnesses can prompt sensibly. Lazy on-demand loading (--tools lazy) is available for token savings
- Stable element refs: accessibility snapshots return a compact text tree with [ref] handles you pass to click, type, scroll and press-key. A ref from an earlier page is reported as stale rather than acting on the wrong element
- firefox_find: locate elements by visible text or accessible name, with landmark context, instead of reading a whole page
- Page navigation and DOM interaction (click, type, scroll, press keys)
- Visible-text reading with paging: long pages are read in 20,000-character chunks
- Attach to existing user tabs: drive open browser tabs in place with explicit agent-initiated attach
- Never disturbs your browsing: agent tabs in a window you already have open are created inactive, so your selected tab is never taken over, and screenshots capture the agent's tab instead of the page you are looking at
- Vision-sized screenshots: the long edge is capped (1568 px by default) so images are not downscaled again by the model, with region zoom and dynamic readiness detection. Optional save to a private temporary file for harnesses that cannot display inline images
- Annotated screenshots with numbered badges that map to element refs
- Structured page analysis via JSON with configurable limits and explicit truncation metadata (faster and cheaper than screenshots)
- Page state diffing to detect changes between interactions
- Cookie and consent banner auto-dismissal (GDPR overlays, CMP platforms)
- Domain allowlists for agent sandboxing
- Flexible element waiting (CSS selector, text content, or URL pattern)
- No hangs on unpainted tabs: animation-frame waits are bounded, so commands settle in a background tab, a minimised window or another workspace
- Network request and console log monitoring
- Multi-agent and multi-harness safety: tab ownership tracking (including navigation), screenshot mutex, and slot reservations
- Focus loops that keep running across agent sessions (state lives in the host process)
- Orphaned tab cleanup with configurable timeout (CLAUDEZILLA_AGENT_TIMEOUT_MS): set it to 0 to switch cleanup off entirely
- Start Firefox on demand: CLAUDEZILLA_AUTOSTART_CMD launches the browser when the host is not reachable (run as an argv vector, never through a shell). Listing tools and shutting down never start the browser
- Two browser profiles side by side: CLAUDEZILLA_IPC_DIR (POSIX only) relocates the IPC socket, auth token and debug log
- Per-operation timeouts (5s-300s, default 150s)
- A portable "claudezilla-browser" Agent Skill that teaches harnesses the efficient workflow (read refs first, screenshot only when needed)

REQUIREMENTS:
- Supported AI harness: Claude Code, Hermes Agent, Pi Agent, omp, Mercury, or any standard MCP client
- Native messaging host and MCP server (one-command installer; --target for Claude Code, omp, Hermes, Pi and Mercury on macOS and Linux; the Windows installer configures Claude Code). Keep the extension and the MCP server on the same version
- Firefox 140+
- macOS, Linux, or Windows 10/11

PRIVACY & SECURITY:
- All communication is local via Unix socket / Windows pipe (no network exposure)
- Zero telemetry: no data collected or transmitted to external servers
- Page content accessed only on explicit tool invocation
- Console capture and JavaScript evaluation are strictly opt-in (off by default in popup settings)
- The installers' MCP allow-list does not pre-approve firefox_evaluate, so the tool is not silently permitted by the harness either
- Screenshot save paths are restricted to the temp and working directories and written owner-only
- Request bodies are never captured, and typed text is not logged
- Per-agent tab ownership prevents cross-session and cross-harness interference
- Community supported
- Open source (MIT license)

Setup & docs: claudezilla.com