Anomaly, the company behind the open-source OpenCode coding agent, has launched a new product called Browse: an MCP server and Chromium extension that let AI coding agents drive your real browser — the one with your tabs, your logins, and your extensions — rather than a sandboxed imitation. The launch reached the front page of Hacker News on November 16, drawing roughly 240 points within hours.

The problem Browse attacks is familiar to anyone who has watched an AI agent flail at a login screen. Most agent browser setups spin up a fresh, sterile profile with no cookies and no sessions, so anything behind authentication fails immediately. Browse takes the opposite approach: agents use the sites you are already signed into. As the product page puts it: "All your logins. No keys, no setup." Browser-using agents are becoming one of the busiest corners of the latest AI developments — Google has been pushing its own version with Gemini's auto-browse in Chrome, and Anomaly is betting developers want their agents inside the browser they already use.

What Browse Actually Does

Browse consists of two pieces: a Chromium extension and an MCP (Model Context Protocol) server that connects the browser to an agent. Once installed, agents can open tabs in the background — grouped and collapsed out of your way — then read pages, click, type, take screenshots, record tabs, and capture network traffic.

Because it attaches to your existing browser, it works with whatever Chromium-based browser you run: Chrome, Edge, Brave, Arc, Helium, Vivaldi, Opera, and Chromium itself. Firefox is not supported at launch — a gap several Hacker News commenters flagged immediately.

Handoffs Keep the Human in the Loop

The design's most distinctive feature is the handoff. When a task reaches a step that genuinely requires the human — a login, a two-factor authentication prompt, a payment — the agent hands the tab back to you, waits, and continues once you click Continue.

There is also a "lend a tab" option: a toolbar button that hands an agent a tab you already have open, useful when you want the agent to work with a page you found yourself. For containment, an allowed-sites feature restricts an agent's tab to only the sites you name — a direct answer to the obvious question of what stops your coding agent from wandering across your entire browsing life.

Works With Any MCP Client

Browse is not limited to OpenCode. Because it exposes itself as an MCP server, it works with any MCP-compatible client — the company lists OpenCode, Claude Code, Cursor, Codex, and Zed, among others. Installation starts with the Chrome extension, which walks through the rest; a CLI can also be installed directly via an install script, npm (`@opencode/browse`), or bun.

How It Compares to Playwright MCP

Hacker News commenters immediately compared Browse to existing options, particularly the Playwright MCP server and the Chrome DevTools MCP server, both of which already let agents control a browser. The practical difference is state: Playwright-style setups typically drive a fresh test browser, while Browse drives your daily browser with your real sessions and extensions. One commenter who had used Playwright MCP with Claude Code and Codex noted it lets you "grant access to pages and sessions that are already open" — and recommended caution all the same, advising separate browser profiles for agents to limit what an agent can reach if something goes wrong.

That caution is not theoretical. Handing an AI agent your authenticated browser concentrates real risk: your sessions, your saved payment methods, and your work accounts all become reachable from a single tool. Anomaly's allowed-sites restriction and human handoffs for payments and 2FA are mitigations, not guarantees, and security-conscious users will want to treat the combination of private data, untrusted web content, and agent-driven exfiltration channels with care.

Early users also reported rough edges. One commenter said the connection has "a habit of disconnecting," requiring a rerun of the pairing step, and pointed to an open OpenCode pull request meant to bring persistent sessions to the project's integrated browser tool — suggesting Browse may eventually be absorbed into the agent itself.

The Pace Question

Not all the reaction was technical. One commenter argued that OpenCode ships "like 3 projects a week," raising concerns about maintenance depth and security review for a tooling ecosystem moving this fast — a single opinion, but a recurring theme whenever the project releases new products at speed. The counterargument visible on the product page is transparency: Browse's source is public, and the MCP approach means it is not locked to one vendor's agent.

Why the Real Browser Matters

Underneath the feature list, the bet is simple: authentication is the bottleneck of agentic computing. Agents that cannot log in cannot do most real work, and approaches that rebuild sessions through APIs or key-based access only cover a fraction of the web. By letting agents inherit the logins you already have — while explicitly routing logins, 2FA, and payments back to the human — Browse is an attempt to get the best of both worlds: full access, with consent gates at the moments that matter most.

Whether developers trust their daily browser to an agent is now the question the market will answer. Given the traction on Hacker News, plenty of them are willing to try.

Stay Ahead of AI

The AI landscape moves fast, and the stories that matter most often break between the headlines. Follow our AI news coverage for reporting you can verify.

Read more AI news →