An AI that lurks in your QQ group until it decides to speak
qq-bridge plugs DeepSeek Harness agents into QQ group chats, where they read along, pick their moments to speak, and route their tool-approval prompts back to your phone.

What it does qq-bridge forwards QQ messages — arriving via SnowLuma’s OneBot v11 WebSocket — into DeepSeek Harness (DSH) agent sessions, and sends agent output back the other way: replies, clarifying questions, even tool-approval requests you can answer from QQ. Each QQ friend or group conversation maps to its own DSH session, all grouped under a single QQ-chat workspace. The agent also gets a deliberately constrained set of MCP tools for QQ itself: check status, read group history, send messages, reply to a specific message.
The interesting bit
The reserved mode is a small sociology experiment: the agent runs a watching/active/probing/leaving state machine to decide when to participate, can emit [SILENT] to deliberately say nothing, and splits its output on spaces so it types like a person rather than a press release. Its successor, reserved2, drops auto-forwarding entirely — the agent polls qq_get_unread_messages and chooses for itself when to speak, wait, or keep lurking.
Key highlights
- Four run modes:
chat(whitelisted safe chat),closed-agent(owner-only, full DSH control from QQ),reserved(simulated group member),reserved2(fully autonomous) - Quote parsing injects who-said-what context, so the agent knows a group member is replying to someone else, not to it
- Two-layer persona system: a built-in simulation prompt for behavior plus editable persona files (
roles/*.md) for tone — roleplay is set by the admin, never by group members - Local web console with a cost dashboard: token spend per group, per friend, per turn, with 24h–30d charts colored by peak/off-peak rates (peak is 2×, Chinese statutory holidays included)
- Slang learning: extracts slang from group chat and researches it through a read-only, SSRF-guarded web-search tool
Caveats
- Version-locked: v0.1.7 targets DSH 0.1.7-rc.2, and the wire protocol changed at DSH 0.1.2 — older DSH installs need older tags of this bridge
- Data leaves the machine by default: search keywords go to Bing, and slang auto-research is on unless you turn it off
- SnowLuma is a proprietary component the bridge won’t install or start for you (its EULA demands written permission for automated deployment), and its SDK ships an ESM bundling bug the bridge patches automatically
Verdict If you already run DeepSeek Harness and SnowLuma, this is remarkably thorough glue — the simulation modes and approve-from-chat design are genuinely thought through. Everyone else can admire it from a distance: it’s married to two specific ecosystems, and to specific versions of them.
Frequently asked
- What is Derpyu520/qq-bridge?
- qq-bridge plugs DeepSeek Harness agents into QQ group chats, where they read along, pick their moments to speak, and route their tool-approval prompts back to your phone.
- Is qq-bridge open source?
- Yes — Derpyu520/qq-bridge is an open-source project tracked on heatdrop.
- What language is qq-bridge written in?
- Derpyu520/qq-bridge is primarily written in JavaScript.
- How popular is qq-bridge?
- Derpyu520/qq-bridge has 3.7k stars on GitHub and is currently accelerating.
- Where can I find qq-bridge?
- Derpyu520/qq-bridge is on GitHub at https://github.com/Derpyu520/qq-bridge.