ctoldigital/axuse

Accessibility-first browser & computer use for AI agents — one MCP server for every harness (Codex, Claude Code, OpenCode, Antigravity). Instruments the DOM/accessibility tree via CDP instead of screenshot loops: diffed snapshots, stable element refs, trusted input, typed errors. Zero dependencies

What it solves

AI agents often need to interact with web pages, but the traditional way — taking screenshots and guessing pixel coordinates — is slow, brittle, and error-prone. axuse solves this by giving agents a structured, text-based view of the page using the browser's accessibility tree and DOM instead of images. That makes automation faster, more reliable, and much cheaper in tokens.

How it works

axuse runs as an MCP (Model Context Protocol) server that plugs into agent harnesses like Codex, Claude Code, OpenCode, or any MCP-capable host. When an agent asks to open a page, axuse returns a compact accessibility snapshot with stable references for each interactive element. Actions like clicking, typing, or scrolling target those references directly. Only changes in the page state are sent back, not full snapshots. It uses Chrome DevTools Protocol (CDP) to send trusted input events, avoiding detection. It also has a server mode for headless Linux machines, an extension bridge to control your own logged-in browser, and desktop control on Windows, macOS, and Linux.

Who it's for

Developers building AI agents that need browser or desktop automation, especially those working with agent harnesses like Codex, Claude Code, OpenCode, or Antigravity. It is also packaged for non-developers: a downloadable file can be double-clicked in Claude Desktop, and an install prompt can be pasted into other agents to set it up automatically.

Highlights

  • Zero npm dependencies (Node.js 22+ only)
  • 28/28 server-scored tasks passed, with median time 17s vs 26s for Codex+extension and 56s for browser-use; uses 4.7M tokens vs 9.8M and 8.8M
  • Diff-based observation: only changed lines are returned, saving tokens and latency
  • Typed errors like REF_NOT_FOUND and STALE_STATE with retry hints, instead of silent misclicks
  • Extension bridge to drive your own daily Chrome/Edge with your logged-in sessions
  • Server mode runs headed Chrome on a virtual display to avoid anti-bot detection (verified against Turnstile)
  • Optional CAPTCHA solver plugins (commercial) with a free reference implementation
  • Desktop control for Windows (verified), macOS (opt-in), and Linux (X11/XWayland)

Related

  • Project
    ZJU-REAL/HugAgentOS

    HugAgentOS is an open‑source, self‑hosted “AgentOS” that lets AI agents retrieve knowledge, run code, and complete real‑world tasks. It uses a domain ontology as a control plane for rule‑based, traceable execution, and offers chat, private RAG, sub‑agents, skill extensions, sandboxed code, multi‑layer memory, automation, and a web UI. Install via a one‑command script (SQLite) or Docker‑Compose (PostgreSQL/Redis, optional Milvus/Neo4j). Community Edition is Apache‑2.0; Enterprise adds multi‑tenant governance.

  • Project
    hydro13/tandem-browser

    A local-first Electron browser that integrates AI agents directly into the runtime, allowing them to automate real SaaS platforms and interact with authenticated web sessions.

  • Project
    zhangfengcdt/memoir

    A high-performance semantic memory system for AI agents that introduces Git-like version control, hierarchical semantic paths, and cryptographic integrity to memory management.

  • Project
    fuxiaoai/tidings-rss

    A curated RSS feed collection (718 feeds, 16 OPML bundles) with strict quality checks, paired with a macOS reader app featuring AI summaries, translations, podcast playback, and FreshRSS sync.