OpenCode Browse is an open source tool that lets an AI agent use the web browser you already have open, with your tabs, your logins and your extensions. It's an MCP server plus a Chromium extension, published on the OpenCode site by Anomaly. The agent opens its own background tabs, reads pages, clicks, types and takes screenshots. When it hits a login, a 2FA prompt or a payment, it hands the tab back to you and keeps going after you click Continue. Version 0.1.0, a rewrite in Rust, shipped on November 16, 2026, after earlier 0.0.x releases from early November. The latest stable release on GitHub was v0.1.5 (November 17) when I checked on November 20.
If you saw "browse" in an agent's tool list, or you're deciding between giving an agent a browser it controls and giving it yours, this guide covers how it works, how to install it, what the safety controls really do, and when another tool fits better.
What is OpenCode Browse?
Most agent browsers start from a blank slate. They launch a clean headless browser, which means the agent has no logins and every site treats it like a stranger. OpenCode Browse flips that. It drives the Chrome, Edge, Brave, Arc, Vivaldi, Opera or Chromium window you use every day, so the agent can open a site you're already signed into, such as your email or an internal dashboard, with no API keys and no extra setup (OpenCode Browse page).
It has two parts, per the project's README on GitHub (anomalyco/browse):
- A small native program that speaks the Model Context Protocol (MCP) to your agent. If MCP is new to you, our guide to MCP servers explains the idea.
- A Chromium extension that runs the agent's commands inside your browser, through Chrome's debugging interface.
Between them sits a relay that listens on your own computer only (127.0.0.1) unless you deliberately open it up. The README says nothing leaves your computer except what your agent sends to its own model provider.
It works with OpenCode, Claude Code, Codex, Cursor, Zed and any other MCP client. The project is MIT licensed. It does not take pull requests, only issues, so don't expect to contribute code.
How do you install it?
The OpenCode page says to install the Chrome extension first and let it walk you through the rest. If you'd rather start from the terminal, the README lists these options:
curl -fsSL https://opencode.ai/browse/install | bash
npm install -g @opencode/browse && browse installOn Windows, the README gives an irm ... | iex line for PowerShell. browse install adds the MCP server to the agents it finds on your computer and pairs your browser. If pairing doesn't happen by itself, browse pair prints a one-time code that expires after 10 minutes. Restart any agent that was already running so it loads the new tools. Then try something small, like asking your agent to open a news site and tell you the top story.
What can an agent do with it?
The README lists the tools. The ones you'll notice first are navigate, snapshot (a compact outline of the page), click, fill, type, screenshot, network (to capture traffic) and recording (to record a tab as video). There's also a handoff tool for asking you to step in, and a tabs tool for listing and closing tabs.
Two details shape how it feels to use:
- Tabs stay out of your way. Agents open background tabs in a collapsed group and only touch tabs they opened or that you lent them. You lend one by clicking the toolbar button on it. Tabs nobody uses for 10 minutes close on their own, unless you change that setting.
- Handoffs. When a human step is needed, the page shows a small card with the site's name. You do the login or the payment yourself, then click Continue. Browse also reports Cloudflare and captcha walls and passes interactive ones to you. The README says it never solves them.
OpenCode itself gets the individual tools. Other MCP clients get three: execute, status and tabs, where execute runs JavaScript that can call all the others.
How safe is it to let an agent use your real browser?
This is the question that matters most, and the README is direct about it. Browse drives a browser you're signed into, so it assumes a web page or an email the agent reads might try to steer it. Here's what the project says it does:
- Every request to the relay needs a token stored in a file only you can read, and each browser pairs once with its own key.
- Clearing cookies, wiping site data, closing the browser and changing site permissions are refused.
- Agents can't open browser internal pages, other extensions or local files.
The options people tend to trust too much are allowedOrigins and readOnly. The README calls them guardrails, not a data boundary. allowedOrigins limits where an agent can navigate, but requests the page makes by itself, like scripts, images and iframes, aren't filtered. readOnly blocks clicks and typing, yet eval can still run JavaScript in the page and read whatever the page can. The README's advice is blunt: for pages you don't want an agent to read, don't open them in browse.
A few more things to know before you turn it on:
- Masking is partial.
fillhides values only for password and one-time-code fields. Page text, screenshots andevalresults come back as they are. - It asks for wide permissions. The extension needs the
debuggerpermission and access to all sites, and your browser shows its "started debugging" bar while an agent is attached. - Updates are automatic by default. The relay checks hourly and installs new releases. You can switch to notify only or off.
- It's early. The version numbers start with 0.1, and the project's first public releases are only a few weeks old (the repository was created on October 27, 2026).
- The remote option isn't encrypted by browse. If your browser is on a different machine, the README says to use a network that is, like Tailscale or WireGuard.
A sensible first rule: use a separate browser profile for agent work, one that isn't signed into your bank or your admin consoles.
OpenCode Browse, Stagehand, Kitesurf or plain Playwright?
They solve different problems, so the best pick depends on whose browser the agent should drive.
| Tool | Whose browser | Best for |
|---|---|---|
| OpenCode Browse | Yours, on your computer, with your logins | An agent that works in sites you're already signed into, with you nearby to hand off |
| Stagehand | A browser your code launches, local or on Browserbase | Developers building browser automations with plain English steps mixed into code |
| Kitesurf | Cloudflare's agent browser, running in Cloudflare Workers | Agents that need a hosted browser without using yours |
| Playwright or Puppeteer | A browser your script launches | Tests and scripts with exact selectors |
The trade is simple. A fresh, hosted browser is safer for untrusted pages because it has no logins to lose. OpenCode Browse is more useful when the work is behind your login, and riskier for the same reason. If you're unsure, start with a throwaway profile and a task you could do by hand in five minutes.
Go deeper
- What is Stagehand: the Browserbase SDK for browser agents.
- What is Kitesurf: Cloudflare's browser built for agents.
- What is an MCP server: how agents get tools like this one.
- AI tools for developers: the 2026 tools, sorted by what they do.
The skill behind it
Tools like this one change every few weeks. What lasts is knowing how to give an agent just enough access, test it on small jobs and read a project's own security notes before you trust it. If you want to build that skill with models and agents at your own pace, see how AI Flex works.
Sources, checked November 20, 2026: the OpenCode Browse page, the anomalyco/browse README, license and releases on GitHub.
