Give your AI agent eyes.
12MB binary. HTTP API. 10x cheaper than screenshots.

🪼 See the web through your agent's eyes

Browser automation tools are designed for humans, not agents. Playwright, Puppeteer, Selenium – they all return massive DOM dumps or screenshots that burn through tokens. And most agent-specific solutions are locked into their own ecosystems.

Seebro takes a different approach: accessibility trees over HTTP.

# Get page content — ~800 tokens instead of 10,000+
curl localhost:9867/text

# Interact with elements using stable refs
curl -X POST localhost:9867/action -d '{"kind":"click","ref":"e5"}'

Bottom line: 5-13x cheaper for typical browsing tasks, and it works with any agent that can make HTTP calls.

docker run -d -p 9867:9867 --security-opt seccomp=unconfined seebro/seebro
curl http://localhost:9867/health

go build -o seebro .
BRIDGE_HEADLESS=true ./seebro

If you're using OpenClaw, just ask your agent to set up Seebro. The skill file has everything it needs.

BRIDGE_HEADLESS=true ./seebro

Chrome runs invisibly. This is what Seebro is optimized and tested for. Best for servers, CI pipelines, and automated workflows.

./seebro

Opens a Chrome window you can watch and interact with. Useful for debugging and manual logins. Note: headed mode has some rough edges – profile management is manual, and some sites behave differently.

Instead of HTML, Seebro returns the browser's accessibility tree – the same structure screen readers use. Each element gets a stable ref (e0, e1, e2...) that you use for interactions.

e0:RootWebArea "Example Domain"
e1:heading "Example Domain"
e2:link "More information..."

Click e2 and the link activates. No coordinate guessing, no CSS selectors that break.

Real measurements from a search results page:

For a 50-page monitoring task:

• Screenshots: ~$0.30
• Full snapshots: ~$0.16
• Seebro /text: ~$0.01

Log in once, stay logged in. Seebro maintains a Chrome profile at ~/.seebro/chrome-profile/ that persists cookies, auth tokens, and open tabs across restarts.

Environment variables control behavior:

Full list in the configuration docs.

Seebro patches common automation fingerprints:

• Removes navigator.webdriver flag
• Spoofs realistic plugins and languages
• Hides CDP artifacts
• Optional: canvas noise, WebGL vendor spoofing, font metric fuzzing

light mode (default) handles most sites. full mode adds fingerprint randomization but can break some pages.

┌─────────────┐                  ┌──────────┐                  ┌─────────┐
│   Agent     │  ── HTTP ───►    │  Seebro  │  ── CDP ───►     │ Chrome  │
│  (any kind) │   JSON req/res   │  (Go)    │   DevTools       │         │
└─────────────┘                  └──────────┘                  └─────────┘

Seebro is a translation layer. Agents speak HTTP, Chrome speaks CDP. Seebro handles the conversion, adds stealth, manages sessions, and compresses output to save tokens.

Seebro gives agents real browser access with your real accounts.

When you log into sites, those sessions persist. Any agent with Seebro access can act as you – read messages, make purchases, access sensitive data.

Essential precautions:

• Set BRIDGE_TOKEN – without it, anyone on your network has access
• Treat ~/.seebro/ as sensitive – it contains your browser profile
• Start with test accounts – don't connect production accounts until you trust your agent
• Firewall appropriately – Seebro binds to all interfaces by default

Think of it as handing someone an unlocked laptop. Powerful when intentional. Dangerous when careless.

Everything else is Go standard library.

• Go 1.25+ (or download a release binary)
• Chrome or Chromium

git clone https://github.com/ChrisFeldmeier/seebro.git
cd seebro
go build -o seebro .
go test ./...

_{Chris Feldmeier · Creator}

Seebro works natively with OpenClaw, the open-source AI assistant. Use it as your agent's browser backend for efficient, low-cost web automation.