All postsSHORT2 min read
OpenCode Browse: Pages May Try to Steer Your Coding Agent
OpenCode Browse lets agents use your signed-in browser. Its README calls site limits and read-only tabs guardrails; eval and webmcp can still change a read-only page.
A web page can try to steer your coding agent. OpenCode Browse's own README says so. It lets agents use your real browser, including your tabs, logins and extensions. That makes its security section worth reading before you connect it.
What Browse does
Browse is an MCP server plus a Chromium extension from Anomaly, the company behind OpenCode. It works with OpenCode, Claude Code, Codex, Cursor and any other MCP client.
According to the product page, agents open background tabs, grouped out of your way. They can read pages, click, type, take screenshots and record tabs.
When a login, 2FA prompt or payment needs you, the agent hands the tab back. It continues once you click Continue.
The November 16 version
Version 0.1.0 came out on November 16, 2026. The release notes describe a Rust rewrite: one small native binary for the CLI, relay and MCP server, with no Node.js or Playwright underneath.
Browse is MIT-licensed and free to install. The repository was created on October 27, so November 16 is the version date.
The catch in the README
You can limit which sites an agent opens and make a tab read-only. The catch: the README calls these guardrails. After a tab opens, readOnly refuses clicks, typing, filling, uploads and navigation commands. But eval and webmcp still run JavaScript in the page: they can read what the page can, make requests as the page and change it. Requests a page makes itself are unfiltered.
The README says its approach “assumes that what an agent reads (a web page, an email) may try to steer it.” Its advice for pages you don't want an agent to read: “don't open them in browse.”
Browse also reports Cloudflare, captcha walls and access-denied pages. The README says it hands interactive challenges to you and “never solves them.”
Try it
On macOS or Linux, the README gives this install command:
curl -fsSL https://opencode.ai/browse/install | bash
Then add it to your agents:
browse install
Read the README's Security section before choosing which pages to open in Browse.
Here's how you get paid
My opinion: I'd sell a release check for small SaaS teams. After each deploy, have an agent with Browse walk signup and checkout, then send screenshots and a tab recording.
The tools include screenshots, screenshot_diff for comparison with a baseline, and recording for saving a tab to video. Those are the pieces I'd use for this service.
The client needs to supply a staging login. A human still handles the 2FA handoff and clicks Continue. I'd keep security testing outside the offer, given the README's stated limits. No income guaranteed.
Daily email
Get each breakdown in your inbox.
One email each evening, right after the Short goes live: the AI tool it covers, the full write-up with sources, and how to get paid for it. Free, and you can unsubscribe anytime.