agent-browser

Built for the agent that uses it, not a human. Snapshots are dense ref-lines, not aria dumps. find filters the cached tree for free. Every action returns a delta (what changed), so you don't burn a turn re-seeing after each click.

Why

The big browser MCP servers tax the agent every step. Measured head-to-head against the two largest browser-automation MCP servers (Playwright MCP and Chrome DevTools MCP), agent-browser costs an order of magnitude fewer tokens:

The gap widens with page complexity — ~4× on a simple flow, ~12–17× on HN/GitHub — because the competitors re-dump the full tree every step while agent-browser is tiered, capped, and delta-based. All three complete the same real tasks; agent-browser just costs less, ships as one binary, and gets the browser closer to a real human. Cost to exist: ~2,363 tokens/turn (15 tool defs + server instructions).

_{Measured 2026-06-19, Windows, headless, over the public internet. Each server driven through its own MCP client; tokens ≈ chars÷4 on the tool-list and snapshot responses. All three servers completed the login.}

Install

Requires Go 1.26+ and Chrome/Chromium (auto-discovered).

go install github.com/dondai1234/agent-browser@latest

Verify the install, and update later with the same command:

agent-browser --version                                  # check the installed version
go install github.com/dondai1234/agent-browser@latest    # re-run to update

Or: tell your agent to install it

Copy this prompt into Claude Code, Cursor, Copilot Chat, Windsurf, or any agent with shell + MCP-config access, and it does the rest:

Install the agent-browser MCP server and connect it to this client:
1. Run:  go install github.com/dondai1234/agent-browser@latest
2. Verify:  agent-browser --version   (expect "agent-browser v1.0.0" or newer)
3. Add a stdio MCP server named "agent-browser": command "agent-browser", args ["mcp"].
4. Confirm it connects, then tell me it's ready.

Connect to your agent

Add it to any MCP client:

{
  "mcpServers": {
    "agent-browser": { "command": "agent-browser", "args": ["mcp"] }
  }
}

Cursor, Claude Code, Claude Desktop, Windsurf, VS Code Copilot, opencode, Hermes Agent, and OpenClaw all work with this shape (VS Code uses "servers" instead of "mcpServers"). Ready-to-paste configs and per-client file paths are in examples/. Claude Code one-liner: claude mcp add agent-browser -- agent-browser mcp.

If your client reports spawn agent-browser ENOENT, it can't find the binary on its PATH — use the absolute path in command: $(go env GOPATH)/bin/agent-browser (append .exe on Windows).

The agent workflow (fewest round-trips)

navigate https://example.com level=summary   →  refs in the same call
find role=button text="Sign in"              →  [r12] button "Sign in"
click r12                                    →  delta (what changed)
fill r3 "user@x.com"                         →  delta
...

Each action returns what changed. You rarely call see after an action — the delta hands you fresh refs.

Tools (15)

navigate · see (minimal/summary/full) · find · read · click · fill · select · scroll · wait · screenshot · eval · tabs · upload · press_key · hover

Every tool's description is hand-crafted to tell the agent exactly what to pass, what it returns, and the gotcha. navigate takes a level so you can get refs in the same call; press_key fires native key events (Enter submits a form); hover triggers CSS :hover; select matches an option's value or visible text; eval covers anything the snapshot can't expose (canvas, computed state, history, cookies, console errors).

Anti-bot / stealth — on by default (--no-stealth to disable)

• Static tells patched: navigator.webdriver false (via --disable-blink-features=AutomationControlled and dropping --enable-automation); userAgentData/plugins/languages/window.chrome/WebGL/hardware spoofed via a pre-page init script. Verified: webdriver=false, plugins=5, languages=en-US,en.
• Real fingerprint: --headless=new (near-real) by default; --headless=false for the real GPU/canvas/timing fingerprint on hard targets.
• Behavioral realism: a jittered smoothstep mouse path before each click; press_key for typed input.
• Proxies + challenge detection: --proxy-server for residential proxies (the biggest IP-reputation lever); navigate/see surface CHALLENGE: on Cloudflare/DataDome/reCAPTCHA/hCaptcha/Turnstile and auto-wait for managed challenges to clear.

Flags

--headless · --user-data-dir (override the profile location) · --no-persist (throwaway profile; by default logins persist at <os config dir>/agent-browser) · --proxy-server · --user-agent · --viewport W,H · --no-stealth · --no-eval (eval is on by default) · --allow-insecure-schemes · --version

Status

v1.0.0 — first public release. 18 live tests (real sites: a saucedemo full purchase + React login, iframes, upload, JS alerts, GitHub/Wikipedia extraction, a CSS :hover menu, Enter-to-submit) plus unit tests. govulncheck clean. CI: ubuntu/windows/macos.

CHANGELOG  ·  Example MCP configs  ·  License: MIT