OpenCode Browse

OpenCode’s new “Browse” tool, which lets AI agents control a user’s existing Chromium-based browser tabs with their logged-in sessions and extensions, is drawing both enthusiasm and alarm. Fans see it as a powerful alternative to built‑in headless browsers and Playwright/CDP setups, enabling richer automation and debugging workflows, while critics argue that OpenCode’s broader security posture, lack of strong permissioning by default, and history of incidents make handing it full browser access risky. The conversation widens into a comparison of agent harnesses, sandboxing strategies, and whether protocols like MCP or direct CDP integrations are the safer long‑term way to let AI interact with real browsers.

Product & Feature Impressions

  • Browse lets agents control a user’s real Chromium-based browser, including existing tabs, logins, and extensions.
  • Several users find it fast, stable, and useful, especially via its CLI and MCP modes; some prefer it to other browser bridges.
  • Others say their existing setups (CDP, Playwright, custom extensions, computer-use tools) already solve the same problem.

Browser Support & Technical Approach

  • Only Chromium-family browsers are supported; Firefox and Safari users feel left out and criticize marketing language like “real browser” when Firefox isn’t supported.
  • Some note Firefox has its own MCP server and say building a Firefox extension with full access is not very hard.
  • Browse is seen as overlapping with Playwright MCP and Chrome DevTools MCP; debate centers on whether a dedicated MCP server is needed when CDP and existing tools exist.

Security, Permissions & Sandboxing

  • Many express serious concern: the extension effectively lets an agent act as the user in authenticated sessions, which is likened to “open doors at a virus lab.”
  • Prompt injection, data exfiltration, and account destruction (e.g., invoices vs. delete-account risk) are recurring worries.
  • There is a broader argument about where security should live:
    • One side: harness-level permissions and auto-review modes are essential; tools that default to “permit all” are unacceptable.
    • Other side: in-harness permission prompts create fatigue and false safety; real protection should come from OS/container sandboxes and tightly scoped tools.

Project Velocity, Stability & Maintenance

  • OpenCode is seen as shipping many projects at high speed; some call this impressive, others see it as “vibecoding” and “throwaway” with poor backward compatibility.
  • OpenCode v2, in particular, is criticized for weakening security defaults, breaking configs, and losing pinned sessions; defenders say you can harden it via config and external sandboxing.
  • There are complaints about mounting GitHub issues, abandoned or under-maintained earlier products, and opaque handling of past security incidents.

Comparisons & Alternatives

  • Users compare this stack with Codex, Claude Code, Pi/oh-my-pi, OpenHands, various harnesses (e.g., Aider, crush, goose, maki, dreamcoder, hax), and Vercel’s agent-browser.
  • Some still prefer OpenCode overall (especially its TUI and tabs); others have switched away due to security and reliability concerns.