Skip to content
New Product

Give Your AI a Real Browser.

Agent Browser is a macOS menu bar app that gives AI assistants full control of a real Chromium browser — locally, privately, with no cloud intermediary.

The local browser automation engine for Claude, Cursor, and CrewAI.

Perfect for OpenClaw and Claude AI. Stress-tested by Claude in April 2026: 9.5/10 →

21-day free trial
No signup. No credit card.
£59£29
Limited launch pricing: 51% off

Or get it free: Agent Browser is included when you buy Agent Mac (£44.99). See pricing →

# How it flows
You (in Claude) → MCP tools → Agent Browser → Chromium

macOS 14 Sonoma or later · 90 MB download

Agent Browser

Your AI Can Read the Web. Now It Can Use It.

Agent Browser runs a local automation engine that exposes 37 powerful MCP tools to any connected AI agent. Ask Claude to research a topic, fill out a form, log into a site, or extract clean content from a page — it handles the browser directly while you stay in the conversation.

Industry Firsts

Why Agent Browser

First

Full Programmatic Multi-Tab Control

The first AI agent browser with automatic tab registration, provenance tracking, and agent-callable tab switching via MCP. Your AI knows which tab spawned which — and can navigate between them on demand.

First

Native Mac Menu Bar AI Browser

The first native Mac menu bar app that packages a full AI agent browser — with GUI, session management, and MCP — in one app. No terminal, no config files, no Docker: one click in Setup downloads the automation browser (Chromium) on first run.

Works With What You Already Use

One click for Claude Desktop, Claude Code and Cursor, with no manual JSON editing. You can add other MCP clients by hand over stdio. CrewAI and scripts use the local HTTP API.

Claude Desktop

one click · MCP stdio

Claude Code

one click · MCP stdio

Cursor

one click · MCP stdio

OpenClaw

manual setup

CrewAI

via local HTTP API

Everything Your AI Needs to Control the Web

Not a scraper. A real browser engine your AI drives step by step, and can check that each step landed.

Dual-Mode Operation

Fast headless Chromium by default, and you can show or hide its window with one call. For sites that want a genuine Chrome, turn on "Use your real Chrome" in Settings (needs Google Chrome installed). It runs a separate profile seeded with a copy of your Chrome cookies, and takes effect after you restart the engine.

Session Persistence

Save and restore cookies and localStorage by name. Your AI stays authenticated across conversations — tested with 169 cookies, 3 origins.

Multi-Tab Awareness

Open, track, and switch between tabs. Full provenance tracking shows which tab spawned which — agents never lose context.

Clean Content Extraction

Fetch any URL as clean markdown, with up to 67% fewer tokens than raw HTML. Extraction runs on your Mac with no API key. Developers can opt in to the Firecrawl cloud service as a fallback for near-empty pages by setting their own key.

Host-Based Execution

Runs on your actual machine — real network access, real browser fingerprints, and access to sites that block sandboxed cloud browsers.

Real-Time Dashboard

Live run monitor at localhost:3210 with action logs, screenshots, and status. See every step your AI takes and why.

ARIA-Mode Snapshots

Token-efficient accessibility tree snapshots help agents find the right elements without processing full HTML.

Local & Private

Your sessions, saved logins, screenshots and run history stay on your Mac, and its services listen only on 127.0.0.1. No cloud browser in the middle: sites see a normal browser on your Mac. Its own calls out are the update check, the licence check and Setup downloads, plus any AI or Firecrawl service you add a key for.

Error Resilience

Invalid or blocked URLs, DNS failures, bad selectors, hidden or covered elements and slow pages come back as structured errors, most with a typed code and a recovery hint, so your AI knows what to do next.

37 MCP Tools. Total Browser Control.

Navigate, click, type, fill, scroll, hover, select, press, wait, screenshot, snapshot, extract, session save/restore, and multi-tab management.

navigate

Go to any URL, with HTTP status reporting.

open

Navigate, wait for the page to settle and return clickable refs — in one call.

act

Click, type, fill, hover, scroll, press — batched and verified.

get_location

Cheap orientation: current URL and title (~20 tokens).

wait_for_settled

Wait for the page to quiesce before reading or acting.

read_network

Inspect captured requests — see the 401/redirect behind a failure.

read_console

Read console output and uncaught JS errors.

set_budget

Daemon-enforced action and time budget that stops runaway loops.

check_backpressure

Detect rate-limit and anti-bot signals before continuing.

set_safe_mode

Confirm-gate irreversible actions for unattended runs.

confirm_action

Approve and execute an action that safe mode gated.

plan_checkpoint

Persist and resume a multi-step plan across restarts.

idempotency_guard

Never re-run an irreversible step after a resume.

site_memory

Per-domain learned site facts, so the agent stops re-discovering each site.

download

Capture a file download to disk, with sha256.

set_dialog_policy

Auto-handle native alert / confirm / prompt dialogs.

scroll_until

Exhaust infinite feeds until a stop condition is met.

fill_form

Fill many fields and verify each one stuck.

macro

Record and replay a verified, parameterized flow.

vault

Log in from a Keychain credential — the secret never reaches the agent.

snapshot

ARIA accessibility (or DOM) snapshot, token-efficient.

screenshot

Capture the page as an image the AI can actually see.

extract_clean

Any URL as clean markdown — ~67% fewer tokens.

extract_structured

On-device structured extraction: markdown plus metadata, or typed fields.

detect_blocker

Spot login walls, 2FA, CAPTCHA and Cloudflare interstitials.

page_report

One-call diagnosis of why a page isn't working.

request_human_assist

Hand control to the human for 2FA/CAPTCHA, then resume.

resume_human_assist

Signal the human has finished, unblocking the run.

get_state

Is the daemon up? Visible or headless, current URL and active guards.

list_runs

Full audit trail of recent operations.

watch

Watch a page on a schedule and get a Mac notification when it changes.

toggle_window

Switch between headless and visible mode.

get_forms

Scan the page for inputs, selects and textareas.

session

Save, restore or import a logged-in session so you sign in once.

list_tabs

List all open tabs with provenance.

switch_tab

Switch tabs by ID. Your own tabs need your OK first.

close_tab

Close a tab by ID. Your own tabs need your OK first.

Benchmarked Performance

Headless mode. Real numbers.

454ms

Avg navigation

across 10 diverse sites

77ms

Fastest nav

cached endpoints

~20

Tokens to orient

get_location instead of a full snapshot

<700ms

8 tabs parallel

opened simultaneously

<3s

90-item scrape

3 pages paginated

20

Selectors

strategies passing

How It Works

Four steps to browser automation.

1

Install & Start

On first run, the Welcome Tour (or Settings → Setup → Install Missing) downloads the automation browser, a private copy of Chromium. Then Start All launches all three services: MCP server, browser engine and dashboard.

2

Auto-Connect

In Settings, Connect AI clients… registers Agent Browser with Claude Desktop, Claude Code and Cursor (whichever you have) in one click, with no manual JSON editing. Restart the client to pick it up.

3

AI Takes Control

Claude calls tools like navigate, snapshot, and act to control the browser directly.

4

Watch Everything

Screenshots, action logs, and results are stored locally and viewable in the real-time dashboard.

Built For You

Power Users

Want Claude to handle web tasks — research, form submissions, price tracking, content extraction — without copy-pasting between windows.

Developers

Building AI agents that need reliable browser automation through a clean MCP interface.

CrewAI Users

Connecting multi-agent pipelines to a live browser via HTTP for complex web workflows.

Cursor Users

Want your AI coding assistant to navigate docs, search issues, and read web content inline.

Technical Specs

Browser engine
Playwright (Chromium) or real Google Chrome via CDP
MCP transport
stdio (Claude Desktop / Claude Code / Cursor) or HTTP (CrewAI / scripts)
Storage
SQLite for run history, filesystem for screenshots
Memory
Built-in per-site memory (learned selectors and macros) in local SQLite; optional long-term memory of autonomous runs in a local Qdrant vector DB. Saved sessions are local files.
Tools
37 MCP tools
Dashboard
Local web dashboard at localhost:3210, opened from the menu bar app with a private access code: auto-refreshing run history, run screenshots and scheduled page watches
Privacy
Runs on your Mac: sessions, saved logins, screenshots and run history stay local. Its own calls out are the update check, the licence check and Setup downloads, plus your AI provider if you turn on autonomous mode.

System Requirements

macOS 14.0 (Sonoma) or later
Apple Silicon or Intel Mac
Node.js bundled — no separate install
Chromium download on first run (one click in Setup)
Google Chrome only for the optional real-Chrome mode
Python 3.13 for CrewAI (optional)
Qdrant (long-term memory) + an Anthropic API key (autonomous mode), both optional

Ready to Give Your AI a Real Browser?

Everything runs on your Mac. No credentials sent to the cloud. Sessions persist between conversations.

21-day free trial — no signup, no credit card.

macOS 14 Sonoma or later · 90 MB download