OpenCode Browse Explained
Let AI Agents Use Your Real Browser (MCP Server + Chromium Extension)
OpenCode Browse is an MCP server and Chromium extension that lets AI agents use your real browser. MIT licensed. Covers install, handoffs, comparisons, risks.
OpenCode Browse is a pair of an MCP server and a Chromium extension that lets AI agents use your own browser: your open tabs, logins and extensions. The official page describes it as letting agents use your real browser. It comes from the OpenCode team (an open-source AI coding agent), and the source code is public at anomalyco/browse on GitHub under the MIT License.
This article organizes what can be confirmed from the opencode.ai product page, the GitHub README and the OpenCode product index (as of November 17, 2026): what it does, how to install it, how to use it, how it differs from other tools, and what to watch out for. The official pages give no release date or Browse-specific pricing, so unconfirmed items are marked as not stated.
What OpenCode Browse is: key points
- What it is: an MCP server plus a Chromium extension that lets an agent use the tabs, logins and extensions of your everyday browser profile
- Browsers: Chrome, Edge, Brave, Arc, Helium, Vivaldi, Opera and Chromium
- Clients: MCP clients such as OpenCode, Claude Code, Cursor, Codex and Zed. The README lists about 30 agents, including Claude Desktop, VS Code (Copilot agent mode), Windsurf, Gemini CLI and Cline
- License: MIT (README). Pricing: not stated on the product page (OpenCode itself is described as 100% free to use)
- Release date: not stated on the official pages. The README says updates are automatic, with stable, beta and dev channels
Browse is not tied to OpenCode. Any MCP-connected agent can use it, so it works as a browser component separated from the agent itself.
What it can do
These are the features described on the product page and in the README.
| Feature | Description (product page / README) |
|---|---|
| Real browser | Use sites you are already signed into, with no API keys or extra setup |
| Background tabs | Agent tabs open in the background in collapsed tab groups |
| Handoffs | For logins, 2FA or payments you do the step yourself, then click Continue and the agent resumes |
| Lend a tab | Click the toolbar button to give the agent a tab you already have open |
| Capture | Screenshots, tab recordings and network traffic |
| Allowed sites | Restrict the agent's tabs to the sites you name |
According to the README the tools are grouped as follows. Page reading: navigate, snapshot, screenshot, screenshot_diff. Input: click, fill, type, press, select, hover, drag, scroll. Page control: wait_for, dialog, upload, eval. Monitoring: network, console, recording. There are also emulate, webmcp, tabs, handoff, demonstrate, settings and status, plus execute for clients other than OpenCode, which runs JavaScript that calls the other tools.
How it works
Per the README, the agent starts browse mcp, a stdio proxy that forwards calls to a relay on 127.0.0.1:19988. The relay runs Chrome DevTools Protocol commands, and the extension executes them inside your browser through chrome.debugger. The extension connects to the relay over a WebSocket; native messaging is not mentioned.
In other words it does not launch a separate automation browser. The extension lives in your daily browser and operations reach it through a local relay. The convenience of reusing your sign-in state and the wide blast radius discussed below are two sides of the same design.
Installation and pricing
The product page says that installing the Chrome extension from the Chrome Web Store is enough, since it guides the remaining setup, and that you should not need to install the CLI directly. If you still want the CLI, the official page lists these commands.
| Environment | Command (product page / README) | |
|---|---|---|
| macOS / Linux | `curl -fsSL https://opencode.ai/browse/install \ | bash` |
| Windows (PowerShell) | `irm https://opencode.ai/browse/install.ps1 \ | iex` |
| npm | npm install -g @opencode/browse && browse install | |
| Bun | bun add -g @opencode/browse && browse install |
The README explains that browse install adds the MCP server to your agents and pairs the extension, using a one-time code and an ECDH key exchange. pnpm, yarn and volta options are also in the README.
Pricing is not stated on the Browse page. OpenCode itself is MIT-licensed open source described as 100% free to use, with a set of free models, though connecting other providers may incur extra cost. The LLM cost on the agent side depends on the client you use.
How to use it
After pairing the extension, you simply use Browse from an MCP client. The agent opens background tabs and works, and when a step needs you, such as a login, it hands the tab back. Once you finish and click Continue, the agent carries on.
browse install rewrites the configuration of supported agents automatically. For stdio MCP clients you configure by hand, register browse as an MCP server per the README: the command is browse and the argument is mcp.
# Install (macOS / Linux, official)
curl -fsSL https://opencode.ai/browse/install | bash
# With npm (official)
npm install -g @opencode/browse && browse install
# The stdio proxy started as the MCP server (README)
browse mcpTypical requests would be "open the signed-in admin page and check this month's numbers" or "find the cause of this tab's errors from the console and network." The official page gives no such examples, so these are usage ideas inferred from the feature list (console, network, snapshot and so on).
How it differs from other browser tools
There are several ways to let an agent drive a browser. The table lists only what could be confirmed from official information and this site's existing articles. Details of other tools change quickly, so check each official documentation before adopting.
| Aspect | OpenCode Browse | Playwright MCP | Claude's Chrome integration |
|---|---|---|---|
| Form | MCP server + Chromium extension | MCP server by Microsoft | Chrome extension for Claude products |
| Browser used | Your everyday browser (Chromium-based: Chrome, Edge, Brave, Arc, etc.) | Launches its own persistent profile by default; --extension connects to a running Chrome / Edge (extension required) | Your everyday Chrome |
| Agents that can connect | Any stdio MCP client (OpenCode, Claude Code, Cursor, Codex, Zed and more) | Any MCP client | Claude products |
| Limiting targets | allowedOrigins and readOnly (README calls them guardrails, not a boundary) | --allowed-origins (README says it is not a security boundary) | Site permissions in the product |
| License | MIT | Apache-2.0 | Product feature (not OSS) |
A related article on this site covers Vercel Labs' agent-desktop, which operates desktop apps through the OS accessibility tree rather than the browser. Browse covers what is inside your browser while agent-desktop covers apps outside it, so their roles do not overlap.
Cautions: security and data handling
Handing your real browser's sign-in state to an agent means the agent can do much of what your accounts can do. These are the points from the README.
- Safeguards documented in the README: the relay requires a token readable only by the user and listens on localhost by default. Pairing uses a one-time code and ECDH. execute runs in a QuickJS sandbox with resource limits. Clearing cookies, closing the browser and changing permissions are blocked
- Allowed origins and read-only are not a boundary: the README states that allowedOrigins and readOnly are guardrails, not a data boundary. Do not rely on them alone to protect secrets
- Page-supplied WebMCP tools are unverified: the README treats tools offered by pages as unverified
- Bot challenges are not bypassed: per the README they are reported, not solved
- Prompt injection: the risk of an agent obeying instructions written in a web page or email remains with a real browser. The official page does not elaborate, so prepare on the operations side
- Use handoffs for logins, 2FA and payments: keep the human in the loop as the product page intends
To try it safely, use a browser profile separate from your work one, and start without signing in to sensitive sites such as banking or admin consoles. Always set allowed sites, and have a human confirm irreversible actions such as purchases, sending and deleting. A similar mindset appears in the article on review tools such as Plannotator, where a human checks the agent's plans and diffs.
Screenshots, tab recordings and network logs can capture personal information or credentials. Decide where captured files are stored and do not share them. The repository does not accept pull requests and asks for issues instead (README).
Good fits and poor fits
| Details | |
|---|---|
| Good fit | Having an agent mostly read internal admin pages or login-required services / checking and debugging a web app under development / sharing one browser-control setup across several MCP clients |
| Poor fit | Keeping sensitive accounts open in the same profile / letting agents purchase, send or delete without human confirmation / operating desktop apps outside the browser |
For using several coding agents from one screen, see our T3 Code guide. Browse can be seen as a component that gives such agents a browser to work with.
What could not be confirmed
- Browse-specific pricing, an official release date and the current version number (not stated on the product page or README)
- The Chrome Web Store listing and the details of requested permissions
- Fine-grained feature differences from other tools (Playwright MCP, agent-browser, Claude's Chrome integration)
Sources
- OpenCode Browse product page: https://opencode.ai/product/browse
- OpenCode product index: https://opencode.ai/product
- GitHub anomalyco/browse (README): https://github.com/anomalyco/browse
- Playwright MCP (GitHub): https://github.com/microsoft/playwright-mcp
What is OpenCode Browse?
A pair of an MCP server and a Chromium extension that lets AI agents use your real browser, including open tabs, logins and extensions. It works with MCP clients such as OpenCode, Claude Code, Cursor, Codex and Zed.
Can agents other than OpenCode use it?
Yes. The product page says it works with MCP clients such as OpenCode, Claude Code, Cursor, Codex and Zed, and the README lists about 30 agents including Claude Desktop, VS Code, Windsurf, Gemini CLI and Cline.
What about pricing and license?
The license is MIT (README). The Browse page states no pricing. OpenCode itself is MIT and described as 100% free to use, though connecting other providers may cost extra.
How do I install it?
The product page says to install the Chrome extension from the Chrome Web Store, which guides the rest. From the CLI you can use curl -fsSL https://opencode.ai/browse/install | bash or npm install -g @opencode/browse && browse install.
What happens with logins and payments?
They are handed off to you. You complete logins, 2FA or payments yourself, then click Continue and the agent resumes.
What security points matter?
Because it uses your real sign-in state the blast radius is large, and the README says allowedOrigins and readOnly are guardrails, not a boundary. Try it in a separate profile, set allowed sites and confirm irreversible actions yourself.
株式会社オブライト
Oflight Inc. is a Tokyo-based IT company providing software development, digital signage, web development, network systems, and IoT/AR/VR technology services.
Related free tools (no sign-up, instant results)
Feel free to contact us
Contact Us