A browser built for agents

MCP-native browser control

browxai

A browser, built for agents. Engine-agnostic across Chromium, Firefox, WebKit, Android, and Safari, with a curated surface on top that an AI agent can actually drive without drowning in tokens.

Get startedTool reference

$ npm install -g browxai

agent · snapshot()

# a page an agent can read, not a DOM dump

form”Sign in”[ref=e3]

textbox”Email”[ref=e4]actionable

textbox”Password”[ref=e5]actionable

button”Continue”[ref=e6]actionable

ActionResultcloses the loop

navigatego to a URL

snapshotread the page as refs

findlocate a target

actclick, fill, verify

Why it exists

browxai is not a wrapper over someone else’s MCP server. It owns its own multi-engine transport so it can own the whole session lifecycle.

// surface

Token-efficient

snapshot() is a scoped, paginated, budgeted accessibility tree. find() returns ranked candidates with evidence, not a guess. Results stay small on purpose.

// safety

Safe by default

Capability-gated tools, an origin allow and block list, confirmation hooks, and a hard anti-wedge deadline on every call. Arbitrary JS, full response bodies, the OS clipboard, and network mocking are all off until you opt in.

// clients

Model-agnostic

Any MCP client over stdio. Use it from Claude, Codex, or anything else that speaks the protocol. It is not tied to one model or one vendor.

// sessions

Sessions and lifecycle

Isolated per-session contexts: persistent, incognito, or attach to an existing Chrome you already control. Headed and headless. Config is driven over MCP, with no out-of-band setup.

Wire it into your agent

Read the threat model