Back to Home

OpenCode Browse Is Cheap to Build, Costly to Trust

The $60 Number Is Real. The Bill Is Not.

OpenCode shipped Browse this week: an MCP server plus a Chromium extension that lets a coding agent drive the browser you are already signed into. Same tabs. Same cookies. Same extensions. The pitch writes itself, because the friction it removes is real. No API keys, no headless sandbox, no re-authenticating into a staging dashboard for the fourth time today.

Then, on Saturday, OpenCode developer Dax Raad posted the numbers everyone is quoting. He had rewritten the browser automation layer in Rust for, by his account, sixty dollars. The memory savings are not marginal. They are one of the few AI-tooling claims this month that survived contact with a merged pull request.

The Rust binary is 4.3 MB, against 53 to 64 MB for the Node version. Memory per Browse connector fell from roughly 190 MB to 5.6 MB after a few calls, which is where the 97% figure comes from. The relay that wires agents to the browser moved from 353 MB after six days of Node to 6.6 MB.

  • 4.3 MB: Rust binary, versus 53 to 64 MB for Node
  • 5.6 MB: memory per connector after a few calls, down from about 190 MB
  • 6.6 MB: relay after serving calls, against 353 MB for the old relay after six days
  • 85%: Browse's Stealth Bench v2 score, against 90.5% for Browser Use Cloud
  • Missing: a Stealth Bench score for the old Node implementation

The architecture is honest work. The relay talks to the existing Chromium extension over WebSockets, uses the Chrome DevTools Protocol for tab operations, sessions, recordings and network captures, and embeds QuickJS so agent-supplied JavaScript runs without a Node runtime. Nothing exotic, nothing hidden.

Now the part the celebratory posts skip. That sixty dollars bought model tokens, not a release. Raad used Claude Opus as the main coding model and DeepSeek as a subagent, and the figure excludes the time he spent checking, testing and fixing the result. It also excludes an ongoing liability: Playwright and Patchright used to absorb browser compatibility, authentication changes and security patches. That work did not vanish. It moved onto his team.

Handing Over a Browser Is Bigger Than It Sounds

Browse's actual feature is not speed. It is identity. The agent inherits authenticated sessions, which is exactly what makes it useful and exactly what makes it different from every sandboxed browser tool that came before it.

The safety defaults deserve credit. Agents get tabs they open themselves or tabs a user explicitly shares. Agent tabs open in collapsed background groups. A login, a two-factor prompt or a payment hands the tab back to the human, who finishes the step and clicks Continue. An allowed-sites control keeps the agent on domains you name, and Browse still carries relay authentication, extension pairing and origin restrictions.

But an allowlist constrains domains, not authority. Your bank, your email and your production console are each a single allowed domain. A tool that can click can also click transfer, delete or revoke. The permission model answers where the agent may go, not what it may do once it arrives with your credentials attached.

The history is not reassuring. In October, attackers broke into an older OpenCode system, exposed API keys used with external providers, and left some Console users facing charges reported at around a thousand dollars, across roughly 1.4% of Console accounts. Separately, researchers documented MCP servers able to execute code during OAuth sign-in on Windows and WSL, and a localhost flaw described as a one-click remote code execution path. Each was patched. The pattern held: the boundary between agent and host keeps being the weak point.

Browse does not invent a new bug class. It widens the blast radius. A prompt-injected page no longer steals a sandboxed cookie. It acts as you, in a tab you lent it, on a site where you are already logged in, while the agent records the network traffic for good measure.

What I Would Want Before Installing It

I am not arguing against Browse. The memory work is genuinely good engineering, and the human-handoff design is better than most competitors shipped in 2026. I am arguing that two memory figures and one benchmark score are not a security review. Five questions, none of which the announcement answers:

  • What permission is granted at install: reading open tabs, or acting inside them?
  • What exactly does the network capture record, and where does that data rest?
  • Can content from a visited page reach the JavaScript eval sandbox, or only the agent?
  • What stops a busy-looping eval, given the pull request says Browse still cannot stop one through the extension?
  • Who fixes Chrome compatibility in six months, now that Playwright is no longer in the loop?

The real problem with the sixty-dollar headline is what it teaches. It presents a rewrite as a cost reduction when it is mostly a cost transfer, from vendor-maintained libraries to in-house code that a small team must now own. That is a defensible trade for a tool this popular. It is not free, and the bill arrives in maintenance hours rather than invoices.

So read the changelog, not the post. Memory fell 97%. Stealth scoring slipped against a hosted competitor. Time zone emulation and the download tool were cut. An unstoppable eval loop survived the rewrite. Browse hands an agent your logged-in browser, and the cheapest part of that sentence is the sixty dollars.

Comments

No comments yet. Be the first to share your thoughts!