Let Your Coding Agent Use Your Real Browser With OpenCode Browse

Here is the problem. You ask your coding agent to test your shop, check a dashboard or fill in a form. It cannot, because it has no browser, or it opens a blank one with no logins. So you do it by hand. Every day. That is your time, and your time costs money.

The old fix was to script a headless browser, log in again inside the script, and fight 2FA. The new fix is shorter: lend the agent the browser you already use. That is what OpenCode Browse does.

What the tool is

OpenCode Browse is two things working together: an MCP server and a Chromium extension. The extension sits in your browser. The MCP server lets an AI client talk to it. According to the product page, the agent then controls tabs with Playwright, takes screenshots, records the screen, and captures network traffic as a HAR file. That last one is gold for debugging, because you get every request your page made.

The agent’s tabs open in the background, in collapsed tab groups, so they do not take over your screen. The product is built by the OpenCode team and the source is on GitHub.

It works with any MCP client. The page names OpenCode, Claude Code, Cursor, Codex and Zed. For browsers it lists Chrome, Edge, Brave, Arc, Helium, Vivaldi, Opera and Chromium.

Pricing

The product page does not state a price. It does not say whether there is a free tier or paid plan, and it does not give a launch date. Because the source is public on GitHub, you can read what it does before you install it. What you will certainly pay for is the AI client you connect, since the agent burns tokens while it clicks. I will show a rough number in the math section.

How to set it up

The page says you should not need the command line. The Chrome Web Store extension guides you through the rest. Here is the path I would follow.

1. Install the OpenCode Browse extension from the Chrome Web Store. If you use Edge, Brave, Arc, Vivaldi or Opera, you can use the same Chromium extension.
2. Follow the steps the extension shows you. It walks you through connecting the MCP server to your client.
3. If you prefer the terminal, the page also lists install methods through curl, PowerShell, npm or bun. Use them only if the guided way fails.
4. Open your coding agent (Claude Code, Cursor, Codex, Zed or OpenCode) and check that the new browser tools show up in its MCP list.
5. Give the agent a small first job in a throwaway tab, for example: open this page, take a screenshot, tell me what you see.
6. Turn on the site allowlist before you do anything real. More on that below.

I am not going to invent exact config lines here. The extension guide shows them for your client, and they can change between versions.

The two features that make it safe enough

Handoffs. When the agent hits a login, a 2FA prompt or a payment, it hands the tab back to you. You do the sensitive step. Then you click Continue and the agent carries on. This is the right design. I would never let an agent type my card number.

Allowlists and lending. You can limit a session to specific sites. You can also lend a single tab with a toolbar button, so the agent works on a page you already have open and nothing else. Use both. An agent with your logged-in browser can reach your email, your bank and your admin panels if you let it.

Real use cases

• Test your own site after a change. Ask the agent to open the product page, add to cart, go to checkout and screenshot each step. You get a visual log without clicking.
• Debug slow pages. Have it load the page and capture the network traffic as a HAR file. Then ask it to find the slowest requests.
• Check a dashboard every morning. Lend a tab, ask for a summary of the numbers, and get it in your editor.
• Record a bug. The agent can record the screen while it reproduces an issue. Send the clip to the developer.
• Fill repetitive forms in tools that have no API, with you taking over for the login step.

Do the math: say you spend 10 hours a month on manual checks and data entry in a browser, and that time is worth €15 an hour. That is €150 a month, €1,800 a year. Now say an agent does the same in 100 runs. On a cheap model like Meta’s Muse Ember, a big run with 1M input and 200K output tokens costs about $0.20, around €0.17. One hundred runs is about €17 a month. Even if runs on a stronger model cost ten times more, you are at about €170, which is near break-even, so pick the model by task. These numbers are my example. Your tasks, and the real price of OpenCode Browse once it is stated, will move them.

Limits and weak spots

• No price, no launch date and no operating system requirements on the page. Check the Chrome Web Store listing and GitHub before you rely on it.
• The only safety controls the page describes are handoffs and allowlists. If you skip the allowlist, the agent can touch any site you are logged in to.
• Browser agents misread pages. They click the wrong thing sometimes. Do not let one run unattended on a live admin panel.
• It is a browser tool, not a scheduler. If you want a job to run every morning with nobody there, you need something else on top.

Alternatives

• Claude in Chrome: Anthropic’s own browsing agent, if you are already deep in the Claude world.
• A Playwright script: more work, but full control and it runs without an agent. Good for tests you repeat daily.
• Make or n8n: no-code automation for tools that do have an API. Cheaper and more stable when an API exists.

Source: OpenCode

II

About the author

Ivan Ivanov

Founder ofHawkValue

If it doesn’t save you time or money, you won’t see it here

• Facts, not hypeEvery price and number is checked against the official source
• Always the real priceYou see what it costs before you click
• I say when it’s weakEven when it’s OpenAI or Google

All articles by Ivan →