Camofox Browser

Point an agent at a real website and two things go wrong. The page arrives as 400 KB of HTML that burns the context window before the agent has read anything useful, and the site’s bot detection turns the request away before that. Camofox Browser is a REST server that addresses both: it wraps Camoufox — a Firefox fork that spoofs fingerprint surfaces at the C++ level rather than patching them from JavaScript — and exposes it through an API shaped for agents.

Run it link

  npx @askjo/camofox-browser        # first run downloads Camoufox, ~300 MB
# -> http://localhost:9377
  

Docker is a make up away; railway.toml and Fly.io deploys are included. Idle memory sits around 40 MB because the browser launches lazily and shuts down when unused, so it can share a $5 VPS with everything else. MIT licensed. Interactive API docs are at /docs, and the OpenAPI spec at /openapi.json.

As an OpenClaw plugin:

  openclaw plugins install @askjo/camofox-browser
  

That gives the agent camofox_create_tab, camofox_snapshot, camofox_click, camofox_type, camofox_navigate, camofox_scroll, camofox_screenshot, camofox_list_tabs, camofox_close_tab, and camofox_import_cookies.

The part that actually matters: snapshots and refs link

The API returns an accessibility snapshot rather than the DOM — roughly 90% smaller than raw HTML — with every interactive element carrying a stable ref (e1, e2, e3). The agent clicks e7; it never writes a CSS selector that breaks when a class name changes.

POST /tabs/:tabId/extract takes a JSON Schema whose properties map to those refs via x-ref, so “get me the title, price, and rating” comes back as structured JSON instead of prose the agent has to re-parse.

Long pages truncate with offset-based pagination, screenshots can ride along with the snapshot as base64 PNG, and there are macros — @google_search, @youtube_search, @amazon_search, @reddit_subreddit, and about ten more — for the searches every agent ends up doing. A /youtube/transcript endpoint pulls captions via yt-dlp with no API key.

Sessions, cookies, and logins link

Each user gets an isolated session, persisted to ~/.camofox/profiles/, so an authenticated state survives a restart. There are two ways to get logged in:

• Cookie import. Export a Netscape-format cookie file, drop it in ~/.camofox/cookies/, and the agent imports it. This is off unless you set CAMOFOX_API_KEY — without the key the server rejects every cookie request with a 403.
• VNC login. Log in visually through noVNC, then export the storage state for the agent to reuse.

  export CAMOFOX_API_KEY="$(openssl rand -hex 32)"
  

Keep that key in your shell profile, systemd unit, or platform secrets — not in openclaw.json, which stores plugin config in plaintext.

Before you deploy it link

• Telemetry is on by default. Anonymised crash and hang reports are filed as GitHub issues, with private domains HMAC-hashed and paths, tokens, and IPs stripped. Turn it off with CAMOFOX_CRASH_REPORT_ENABLED=false if any of that is a problem for you.
• Imported cookies are live credentials on disk. A session file that logs into your LinkedIn is worth as much as the password. Treat that directory accordingly, and give the agent an account you’re willing to lose.
• Anti-detection is not authorisation. Getting past a fingerprint check says nothing about whether you’re allowed to be there. Terms of service, rate limits, robots.txt, and the law where you and the site live all still apply, and “the browser could” is not a defence any of them accept. Automating your own account, a site you operate, or public data you’re permitted to collect is the ground this belongs on.
• If there’s an API, use the API. A browser is the fallback for sites that don’t offer one — slower, more fragile, and more likely to break next Tuesday.

Where it fits link

Camofox is infrastructure, not an agent: it gives one hands on the open web, and leaves the judgement about where to put them to you.

Next link

The same skills-and-agents pattern, pointed at markets → Vibe-Trading