What it does

WeClaw is a Go binary that logs into WeChat via QR code, then sits between your WeChat messages and whichever AI agents you have installed—Claude, Codex, Gemini, Kimi, Cursor, etc. You text it, it texts the agent, the agent replies, and you get the reply back in WeChat. It also handles images, files, voice messages (via WeChat’s own speech-to-text), and can proactively message users through a CLI or local HTTP API.

The interesting bit

The project speaks three different “dialects” to agents: ACP (JSON-RPC over stdio, keeping a long-running process warm), CLI (spawning a fresh process per message), or plain HTTP OpenAI-compatible endpoints. It auto-detects which mode is available and prefers the faster ACP. That’s the kind of pragmatic plumbing that makes the difference between “works on my machine” and “actually works.”

Key highlights

• Chat commands inside WeChat: /codex, /cc, /cwd, /new, etc.—including custom aliases you define in config
• Markdown from agents gets stripped down to plain text for WeChat; image URLs are downloaded, encrypted (AES-128-ECB), and pushed to WeChat’s CDN
• Proactive messaging via CLI (weclaw send) or HTTP API on 127.0.0.1:18011
• Background daemon with launchd/systemd service files included
• One-line install script, weclaw update for self-updates, cross-platform releases via GitHub Actions

Caveats

• Docker image ships only WeClaw itself; ACP/CLI agents need their binaries mounted or a custom image built
• README carries a “personal learning only, not for commercial use” disclaimer, inspired by a Tencent WeChat package—legal footing is unclear
• Some agents need --dangerously-skip-permissions or similar flags to run non-interactively, which the README warns against lightly

Verdict

Worth a look if you live in WeChat and want your AI agents to live there too. Skip it if you need a polished, commercially backed integration or if the QR-code-login dance already exhausts you.