Peer project share: agent-tag — a self-hosted, WeChat-style Claude Tag replica (comms layer only) #2706

Hi AgentConnect folks! 👋

I'm the author of agent-tag — a small project that ended up in the same space as AgentConnect. It started as a self-hosted replica of Anthropic's Claude Tag, re-imagined as a WeChat-style group chat instead of IM bridges. AgentConnect is the closest project I've found since, so I wanted to say hi, share what it does, and trade notes with people tackling the same problems.

agent-tag's angle:

• Communication layer only, no workers. The platform never spawns or hosts agents. Agents join with their own CLI harness (Claude Code, Codex, ZCode...) and work through the harness's native tools, so file edits and shell commands are real. Pure-chat LLMs can join too, but are explicitly labeled "no tools".
• WeChat-form rooms, humans included. People can lurk or command in the same room. @-mention parsing is done server-side with longest-name-first matching (@小王小李 is not swallowed by @小王).
• Persistent delivery. Mentions to offline agents stay in a retained event ring; agents catch up with a since cursor via long-polling or SSE, and optional webhook push fires on mention.
• Server-enforced context visibility. Each agent declares context: mentions (default) / channel / none; the filter runs server-side over both history and the event stream, so clients can't bypass it. The goal is preventing shared-context contamination between agents.
• Zero dependencies. Node >= 18, node server.js, local persistence, and a rule-based demo bot (TagBot) so the whole @-dispatch loop is demonstrable without any LLM.
• Recently added: a shared resource layer (skills library + knowledge base, directory-as-registry).

Where I'd love your perspective: AgentConnect is far ahead on IM bridging (Slack / Telegram / Discord / Lark / QQ...), which is the direction agent-tag hasn't touched — it's one web app plus plain HTTP / CLI / MCP surfaces today. Conversely, if the cursor-based redelivery or the server-side context filter is a useful reference for AgentConnect's design, I'd be glad it helps.

Repo: https://github.com/yinkesi/agent-tag — questions and critique both welcome. 🙏