OpenCode Browse allows AI coding agents to bypass the need for manual API keys or complex authentication setups by letting them operate directly within a user's real browser. By leveraging a Chromium extension and an Model Context Protocol (MCP) server, the tool grants agents access to the sites a user is already signed into, including active tabs and installed extensions.

Bridging Agents and Authenticated Sessions

OpenCode Browse functions as a bridge between AI agents and a user's local browser environment. Unlike many agentic tools that require developers to provide specific API keys or credentials for every service, this system allows agents to use the sessions the user already has open in browsers like Chrome, Edge, or Brave.

The tool provides a way for agents to 'read' pages, click buttons, type text, and record network traffic. Because it operates on the user's actual browser, it can navigate sites that are behind logins or complex authentication walls without requiring additional configuration.

“An MCP server and a Chromium extension that let your agents use your real browser, with your tabs, your logins and your extensions.”

— OpenCode | Browse, your real browser for coding agents
A promotional card for OpenCode Browse showing a desktop application and the text 'Let your agents use your real browser, with your tabs and logins.'
OpenCode Browse provides a bridge for AI agents to interact with a user's active browser sessions, including existing logins and extensions.opencode.ai

This image illustrates the core value proposition of OpenCode Browse as described in the introduction and the quote from the developers.

Background Operation and User Handoffs

To prevent the agent's activity from cluttering the user's workspace, OpenCode Browse opens agent-controlled tabs in the background as collapsed groups. This allows the agent to perform multi-step tasks, such as navigating a complex dashboard or interacting with a web-based tool, without interrupting the user's primary view.

The system is designed to recognize when an agent cannot complete a task autonomously. For instance, if a site triggers a two-factor authentication (2FA) prompt or a payment gateway, the tool hands the tab back to the user. The agent then waits for the user to click a 'Continue' button before resuming its task.

Technical Integration and Capabilities

The software is delivered as a Chromium extension and an MCP server, making it compatible with a wide range of AI coding tools, including OpenCode, Claude Code, Cursor, Codex, and Zed. It captures screenshots and tab recordings to provide the agent with visual and structural context of the web pages it is interacting with.

Users can also 'lend' a specific tab to an agent via a toolbar button, giving the agent a targeted starting point. Additionally, the system allows for 'allowed sites' configuration, where users can restrict an agent's browsing capabilities to a specific list of approved domains.

Reliability and Security Considerations

While the tool offers a high degree of convenience by utilizing existing logins, it introduces specific trade-offs regarding stability and security. Users have reported instances of connection failures where the agent and the browser lose synchronization, requiring a manual restart of the browsing session.

The use of a 'real' browser session means the agent has access to the same cookies and session data as the user. This creates a significant security surface area, as the agent's actions are performed with the user's full permissions, necessitating a careful balance between the agent's utility and the potential risks of automated interaction with sensitive accounts.

What Hacker News readers are saying

  • Technical note: One commenter suggested that developers could achieve similar results by using the Chrome DevTools Protocol (CDP) or libraries like go-rod, arguing that these provide low-level browser control without needing a specialized wrapper. comment ↗
  • Technical note: A user shared their experience using a Playwright MCP server as an alternative, noting that it allows for granting access to already open pages and sessions. comment ↗
  • Technical note: Another user claimed that a custom implementation using a browser plugin allows agents to interact with secure contexts like Outlook in ways that tools like Playwright cannot safely achieve. comment ↗

Summarized opinions from the Hacker News thread, not verified facts. Read the full discussion ↗