Agent browsers are web browsers or browser engines designed to be driven by AI agents rather than humans. Instead of a graphical interface, they expose programmatic surfaces — HTTP APIs, MCP servers, CDP endpoints — so an autonomous agent can fetch pages as clean text (typically markdown), render JavaScript-heavy SPAs, take screenshots as visual input, run aggregated searches, and drive multi-step interactions such as logins and pagination.
Common capabilities in this category:
Read: HTML-to-markdown conversion and structured data extraction from JS-rendered pages See: screenshot rendering of the post-JS DOM as vision input for multimodal models Find: meta-search across multiple engines with result merging and deduplication Act: persistent sessions with indexed interaction (click/type by element index), cookie import/export for login flows Evade: real TLS fingerprints and challenge handling for anti-bot protected sitesImplementations range from wrappers around full Chromium (Puppeteer/Playwright-based) to standalone engine binaries that embed their own JavaScript runtime and renderer. The Model Context Protocol has become a common control plane for exposing agent browsers to tools like Claude Code and Cursor.