unbrowse-ai/unbrowse

What it solves

Unbrowse provides a faster, more reliable way for AI agents to interact with the web. Instead of relying solely on browser automation (like Playwright), which is slow and resource-intensive, Unbrowse learns the first-party API routes of a website to execute tasks directly via API calls, bypassing the need to render pages or parse the DOM.

How it works

Unbrowse acts as an action layer. When an agent requests a task, Unbrowse checks its cache for a known API route. If it's a "cache hit," it executes the request directly. If it's a "cache miss," it opens a real browser to observe the working request and record the route for future use. The browser remains available for complex flows like logins, CAPTCHAs, and consent forms.

Who it’s for

Developers building AI agents that need to perform repetitive web tasks efficiently, as well as agents that can be integrated via CLI, TypeScript SDK, or Model Context Protocol (MCP).

Highlights

• Performance Boost: Achieves significant speedups (up to 5.4x median) over traditional browser automation.
• Hybrid Approach: Combines direct API execution with a browser fallback for UI-only flows.
• Local Credentials: Keeps site sessions and credentials on the local machine.
• Shared Learning: Allows sanitized route metadata to be published so other agents can avoid redundant discovery.
• Flexible Integration: Supports a CLI, TypeScript SDK, and MCP compatibility.