Octoweb

What octoweb does

octoweb MCP server lets MCP-compatible AI clients control the Octoweb browser on macOS. The browser is built with WebKit and Rust, and the MCP interface exposes browser automation functions through a local HTTP endpoint. Agents can navigate pages, manage tabs, interact with page elements, capture screenshots, extract content, and work with selected mail, calendar, and task features.

The server is part of the Octoweb browser rather than a separate standalone browser driver. Opening Octoweb provides the browser environment that the MCP tools operate on, including its workspaces, tabs, page sessions, and account-scoped productivity features.

How it works

Octoweb listens for MCP requests at http://localhost:3434/mcp. An MCP client connects to that endpoint and sends tool calls to the running browser. The browser performs the requested action and returns the result to the client.

Browser state is organized through Octoweb’s workspaces. Each workspace has separate tabs, history, AI sessions, and a dedicated WebKit data store, keeping cookies, local storage, and cache isolated between workspaces. Private tabs use in-memory cookies and storage and are discarded when closed.

Setup and configuration

Install and open the macOS application from the Octoweb releases page. The embedded MCP server is available from the browser at http://localhost:3434/mcp; configure an MCP client to use that address.

The browser itself does not require a separate MCP server process or environment variable. Octoweb’s built-in AI sidebar is a separate feature: it uses the local octomind agent server and can require a provider API key such as OPENROUTER_API_KEY, OPENAI_API_KEY, or ANTHROPIC_API_KEY. Those settings are for the sidebar and are not needed for the browser-control endpoint.

Tools and capabilities

The MCP toolset includes capabilities for:

• Opening URLs and navigating browser history.
• Listing, switching, creating, closing, and otherwise managing tabs.
• Clicking, typing, selecting, and interacting with pages.
• Taking viewport and full-page screenshots.
• Extracting page content and searching stored page text.
• Reading and searching mail, with support for the accounts configured in Octoweb.
• Reading calendars and drafting calendar events.
• Listing and adding tasks from the browser’s workspace notes.

Octoweb also provides browser features that agents can work alongside, including isolated workspaces, private tabs, page memory, session restoration, content blocking, extensions, and a terminal panel. The MCP interface is intended for AI clients such as Claude Desktop to drive the browser directly.

Limitations and notes

octoweb is macOS-only and uses WebKit rather than Chromium or Electron. The MCP endpoint depends on the Octoweb application running locally. Account-related mail and calendar operations depend on the accounts configured inside the browser and are scoped per workspace.

The project describes the MCP interface as exposing 44 tools, covering browser control plus mail, calendar, and task operations. The AI sidebar and background learning features involve additional local agent behavior and should be treated separately from the browser-control MCP connection.