OpenCode Browse: Give AI Coding Agents Your Signed-In Browser OpenCode released Browse 0.1.0, an MCP server plus Chromium extension that lets AI coding agents operate inside a developer's already signed-in browser through local background tab groups, with no cloud relay or API keys. The 0.1.0 release rewrote the server in Rust, cutting memory use from roughly 190 MB to 5.6 MB per connector (a 25x reduction) and shrinking the binary from 53–64 MB to 4.3 MB, while dropping Playwright as a runtime dependency. Browse pauses and hands control back to the user at login, 2FA and payment steps, blocks page.bringToFront() so agents cannot steal focus, and auto-closes idle agent tabs after 10 minutes unless a script is running. Every developer who has handed a task to an AI coding agent has watched it stall at a login screen. Headless browsers start cold — no cookies, no sessions, no 2FA tokens — and anything behind authentication is simply off the table. OpenCode’s new Browse tool https://opencode.ai/product/browse rethinks the architecture: rather than giving your agent a sandboxed browser it has to authenticate from scratch, Browse plugs into the browser you already have open and lets the agent operate in background tab groups, using your existing sessions. What Browse Actually Does Browse is two things: an MCP server running locally and a Chromium extension installed in your browser. The two communicate over a local socket — no cloud relay, no API keys, no traffic leaving your machine. Your coding agent sends commands through the Model Context Protocol https://modelcontextprotocol.io , the MCP server translates them, and the Chromium extension executes Playwright-level browser control against your real, signed-in profile. What that unlocks: the agent can navigate, click, type, and take screenshots on any site you’re currently logged into. That means internal staging environments with SSO, admin panels, third-party API consoles, analytics dashboards — workflows that have required manual steps since agents became useful. The Handoff Model Is What Makes This Safe Enough to Use The obvious concern: if the agent has your browser, the agent has your credentials. Browse handles this with a handoff model. When the agent hits a login prompt, a 2FA screen, or a payment step, it pauses and passes control back to you. You complete the sensitive action, click Continue, and the agent picks up where it left off. Agent tabs open in collapsed background groups so they don’t interrupt your work. The protocol blocks page.bringToFront , so an agent cannot yank focus from whatever you’re doing. Idle agent tabs auto-close after 10 minutes — but never while a script is actively running on them. Browse also includes a demonstrate mode : you perform a workflow while Browse records your steps, giving the agent a template to replay and adapt. A lend-a-tab button in the toolbar lets you hand the agent a specific tab you already have open, rather than waiting for it to navigate there. Browse 0.1.0: The Rust Rewrite Changed the Production Math Browse launched alongside a 0.1.0 release that rewrote the entire server in Rust. The original Node implementation worked but consumed roughly 190 MB per connector — enough to notice when running alongside an active coding session. The Rust version dropped that to 5.6 MB, a 25x reduction. The binary shrank from 53–64 MB to 4.3 MB. Playwright as a runtime dependency is gone entirely; browser operations are now exposed directly to the agent. At 5.6 MB, Browse becomes a background tool you forget is running. That changes the calculus on whether to leave it on permanently — the answer is now probably yes. The New Stack covered the rewrite https://thenewstack.io/opencode-browse-rust-rewrite/ , noting that OpenCode’s Dax Raad rebuilt the server in Rust for approximately $60. Installing Browse One command handles the server installation: curl -fsSL https://opencode.ai/browse/install | bash After that: open chrome://extensions in your Chromium-based browser, enable Developer mode, load unpacked from the browse/dist folder, and restart your MCP client. Browse works with Chrome, Edge, Brave, Arc, Vivaldi, Opera, and any other Chromium-based browser. On the agent side, it’s compatible with OpenCode, Claude Code, Cursor, Codex, Zed, and any MCP-compatible client. If your coding tool speaks MCP, Browse works with it. The source is on GitHub https://github.com/anomalyco/browse for anyone who wants to audit or self-host. Browse vs. Playwright MCP: Different Jobs Playwright MCP is still the right choice for repeatable, stateless automation — CI pipelines, test suites, anything that needs a clean, deterministic environment. Browse is for the other half: workflows that require your identity. If your agent needs to check a staging dashboard, file a Jira ticket using your account, and post a Slack update — that’s Browse. If it needs to run a regression test suite against a URL — that’s Playwright. The DEV.to community has a detailed comparison https://dev.to/hassann/opencode-browse-vs-playwright-mcp-vs-chrome-devtools-mcp-which-browser-mcp-should-you-use-2360 of all three browser MCP options if you need to go deeper. The Security Caveat You Should Know OpenCode is explicit about one thing: Browse permissions are user-experience guardrails, not OS-level containment. When an agent runs code through Browse, it runs with your browser’s privileges. The practical guidance: treat it the same way you’d treat running a shell script as your own user. Use the site allowlist feature on shared machines to restrict which domains an agent can touch. Browse is not a sandbox. It’s a delegation layer — and it’s honest about that distinction. For most developer workflows on a personal machine, that’s a reasonable trade-off for the automation you get in return.