Brisk

Give your AI a real browser to use.

Brisk lets an AI agent control Chrome the way you do.

It can open pages, click buttons, type into forms, scroll, take screenshots, read page content, and pull data out of websites. The nice part: it can do this inside the Chrome profile you already use.

So if you are logged in to a site, Brisk can work with that same session. No weird empty browser. No starting from scratch every time.

Think of Brisk as the hands and eyes for Codex, Cursor, Claude, Cline, Continue, ChatGPT Desktop, or any tool that speaks MCP.

The AI still does the thinking.

Brisk gives it the browser.

What Brisk Can Do Today

Brisk V0.1.0 is a working browser-control layer.

It is not a full AI assistant yet. It does not call an LLM by itself. It connects an AI client to a real browser and gives that AI 37 practical tools.

With those tools, an agent can:

• open a new tab
• go to a website
• switch between tabs
• close tabs
• see the current page URL and title
• take screenshots
• inspect the page
• run JavaScript on the page
• read console errors
• click
• type
• press keys
• scroll
• hover
• fill inputs
• upload files
• choose dropdown options
• wait for pages to load
• wait for elements to appear
• make simple HTTP requests
• call raw Chrome DevTools commands when needed
• check the browser connection
• restart the connection
• notice browser dialogs
• read and write site-specific skills
• record failures so future agents can learn from them

That is the core idea.

Instead of asking an AI to guess what is on a page, Brisk lets it look. Instead of asking an AI to imagine a click, Brisk lets it click.

The Best Part: It Can Use Your Real Browser

Most browser automation tools open a clean test browser.

That is useful for tests, but annoying for real work. You lose your logged-in sessions, cookies, extensions, and open tabs.

Brisk's recommended path is different.

It connects to the Chrome, Edge, Brave, or Chromium profile you already use.

That means an agent can help with real browsing tasks like:

• open a product page and extract the price
• summarize a page you already have open
• inspect a web app after you log in
• click through a workflow while you watch
• grab data from a page that needs your normal browser session

There is also a clean isolated-browser mode for tests and unattended jobs. You choose.

Quick Start

1. Install

npm install -g @brisk/cli

Check it installed:

brisk --version

You should see:

0.1.0

2. Turn On Browser Access

Open this in your normal Chrome:

chrome://inspect/#remote-debugging

Tick:

Allow remote debugging for this browser instance

Chrome may show an "Allow remote debugging?" popup the first time Brisk connects. Click Allow.

This is the only slightly strange setup step. Browser automation is still browser automation. Chrome wants your permission.

3. Check Everything

brisk doctor

doctor checks Node, Chrome, the browser connection, and the local Brisk daemon.

If Chrome is open but Brisk cannot connect, doctor tells you what to do and tries to open the Chrome settings page for you.

4. Start Brisk

For most AI apps:

brisk serve

That starts Brisk over stdio, which is the default MCP mode for desktop clients.

For HTTP:

brisk serve --transport http --port 9100

Then connect your MCP client to:

http://127.0.0.1:9100/mcp

Brisk only binds HTTP to localhost by default. That is intentional. V0.1.0 has no authentication, so exposing it to a network is something you should only do on a trusted setup.

Connect It To An AI Client

Brisk is an MCP server.

Any MCP-capable client can use it.

Claude Desktop

Add this to your Claude Desktop config:

{
  "mcpServers": {
    "brisk": {
      "command": "brisk",
      "args": ["serve"]
    }
  }
}

Cursor

Use either stdio:

brisk serve

Or HTTP:

http://127.0.0.1:9100/mcp

ChatGPT Desktop

Add an MCP server with this command:

brisk serve

Generic MCP Client

Brisk supports standard MCP tools and resources.

It exposes:

• tools/list
• tools/call
• resources/list
• resources/read

Example: Ask An Agent To Use The Browser

Once Brisk is connected to your AI client, ask something like:

Open this Amazon product page and extract the title, price, rating,
review count, and availability.

Behind the scenes, the agent can:

1. open a new tab
2. wait for the page to load
3. read the page
4. run a small extraction script
5. return the data

Brisk already includes an Amazon product extraction skill with selectors for title, price, rating, reviews, availability, brand, ASIN, and bullet points.

Important: if Amazon shows a CAPTCHA, Brisk tells the agent to stop and tell you. It does not try to bypass it.

Way 1 And Way 2

Brisk has two browser modes.

Way 1: Your Everyday Browser

This is the recommended mode.

Use it when you want the agent to work with your current browser:

• your logged-in sessions
• your cookies
• your extensions
• your open tabs

Run:

brisk doctor
brisk serve

Way 2: A Fresh Browser For Automation

Use this when you want a clean browser that Brisk owns.

Good for tests, CI, and unattended jobs.

Run:

brisk chrome --port 9222

Then in another terminal:

brisk serve --cdp-port 9222

For headless mode:

brisk chrome --port 9222 --headless

What Makes Brisk Different

It is thin.

Brisk does not try to be a giant automation framework. It does not hide the browser behind a big abstraction. It gives the AI simple, direct tools.

It is local.

Your browser runs on your machine. Brisk talks to it over Chrome's own debugging interface.

It speaks MCP.

That means many AI tools can use it without Brisk needing a special plugin for each one.

It can learn.

When an agent hits a weird website behavior, it can record the failure and write a small skill. Next time, another agent can read that skill instead of falling into the same hole.

What Brisk Is Not Yet

Brisk V0.1.0 is the browser-control foundation.

It is not yet:

• a full AI agent
• an LLM provider
• a browser fork
• a cloud browser service
• a replacement for Playwright or Selenium

It is closer to a very sharp remote control for a real browser.

That is enough to make a lot of useful agent work possible.

Project Layout

brisk/
├── packages/
│   ├── brisk-types     Shared TypeScript types
│   ├── brisk-ipc       Local daemon communication
│   ├── brisk-core      Chrome connection, sessions, browser helpers
│   ├── brisk-skills    Skill store and failure log
│   ├── brisk-mcp       MCP tools and transports
│   └── brisk-cli       brisk doctor / serve / chrome / daemon
├── agent-workspace/    Skills and failure records the agent can edit
├── interaction-skills/ Browser operation notes exposed as MCP resources
├── references/         Upstream projects used for study
└── docs/               Install, protocol, architecture, and roadmap docs

Useful Docs

Doc	What It Is For
Install Guide	Set up Brisk and connect AI clients
Architecture	How the pieces fit together
MCP Protocol	Tool list and MCP wire details
Chromium Embedding	Future browser-fork plan
V0.1.0 Plan	Original detailed build plan

Tech Stack

Layer	Choice
Runtime	Node.js 22+
Language	TypeScript
Browser link	Chrome DevTools Protocol
AI link	Model Context Protocol
HTTP server	Hono
Skill DB	better-sqlite3 with FTS5
Package manager	pnpm

Roadmap

• V0.1.0: local browser control, 37 MCP tools, real-browser attach, stdio and HTTP transports, skill storage.
• V0.1.1: richer skill bundle, profile sync experiments, more remote browser work.
• V0.2.x: an agent SDK on top of Brisk.
• V0.3.x: deeper Chromium integration.
• V1.0: a full Brisk Browser distribution.

License

Apache-2.0.

The projects under references/ are kept for study only.

One Sentence

Brisk turns your real browser into something an AI can see, touch, and learn from.