From 9831594d2cee00ce92c6bd6e3ea4e278587a9e85 Mon Sep 17 00:00:00 2001 From: yinnho <163239465+yinnho@users.noreply.github.com> Date: Tue, 25 Aug 2026 10:52:25 +0800 Subject: [PATCH] Add curated topic definition for agent-browser --- topics/agent-browser/index.md | 21 +++++++++++++++++++++ 1 file changed, 21 insertions(+) create mode 100644 topics/agent-browser/index.md diff --git a/topics/agent-browser/index.md b/topics/agent-browser/index.md new file mode 100644 index 00000000000..f29506f78a3 --- /dev/null +++ b/topics/agent-browser/index.md @@ -0,0 +1,21 @@ +--- +display_name: Agent Browser +topic: agent-browser +aliases: browser-for-agents, ai-browser, agent-web-browser +related: mcp, browser-automation, web-automation, headless-browser, ai-agents +short_description: Browsers and browser engines built for AI agents to fetch, render, search, and act on the web programmatically. +created_by: Browser-use, AginxBrowser, and the agent tooling community +released: 2024 +--- + +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 sites + +Implementations 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](https://github.com/topics/mcp) has become a common control plane for exposing agent browsers to tools like Claude Code and Cursor.