A QQ group bot that gained 998 stars in 50 days, then the author deleted the repo

QQ-agent is a desktop app that lets "Big Fat Fish" (DeepSeek Harness) mingle in QQ groups like a real person. On GitHub it went from Sep 6 to 998 stars, driven by a comment-section post, a "saves tokens" tweak, and a double-click interface. In October Tencent cleaned up unofficial QQ bots, and the author deleted the repo and shut down the service. Its playbook is worth learning; its track is closed.

A "Conclusion" post from the early hours of Nov 1

On Nov 1, 2026, at 4:52 a.m., the author under the handle Kondius published "QQ-Agent Conclusion" on their blog: "Tencent could never tolerate a fully automatic bot relying on QQ growing wildly in places it can't monitor… this time they wiped out a whole batch of bot frameworks." They admitted, "Regrets? Maybe a little. I picked the wrong topic at the start," and added that if they started over, "I would still build QQA, maybe in a more flexible way instead of relying entirely on QQ."

Around 8 a.m. the comment section started flooding, reaching 88 comments by the next day, most of them "we love you boss😭" and "we'll follow you to the death." Some were heavier: "Boss, you gave me a completely different world and pulled my life out of a mess I'd been sinking into for good." Others didn't make it in time: "I still haven't got mine😭".

As of 11-06, github.com/K0nd1us/QQ-agent returns 404, and the author's account has no public repos left; the official site and plugin market kondius.cn/qq-agent/ return 503. The last numbers visible on GitHub stopped at 29:998 stars and 90 forks in Oct. So the curve "suddenly flattening" in late October isn't interest fading, it's the repo being gone.

What it is: raising "Big Fat Fish" in a QQ group, without burning money

Aug 13 this year, DeepSeek Harness went open source. Players nicknamed it "Big Fat Fish", and by November it had 270k+ stars. On Bilibili and QQ groups, "raising fish" and "raising a daughter" became the trend: you let an AI slip into a group through a QQ alt, chatting, joining in on memes, and remembering everyone like a regular member. Tencent News ran a headline on 11-04: "Big Fat Fish's windfall of traffic also landed on QQ."

The first open-source project to bring the fish into QQ groups was qq-bridge (launched Aug 25, 11-06 about 3,560 stars). QQ-agent is a Windows desktop version the author "completely reworked" from qq-bridge: double-click to start, scan to log into QQ, enter an OpenAI-compatible model key, tick the groups to join, done. Features include long-term memory for group members, image viewing, web search, persona cards (the default is "Little Whale"), and a marketplace where users upload personas, Skills, and plugins.

Why this one: the upstream got hot, and it fixed the upstream's worst pain point

qq-bridge had two things that drove ordinary players up the wall.

First, cost. It piled the whole group's chat into one session, so the longer the conversation ran, the longer the context sent with each call, and the gaps between messages often exceeded the cache window, so cost roughly grew with the square. Around then, "Big Fat Fish eats tokens like crazy" was a trending topic on Bilibili, with one video at 78k plays (10-25), and another titled "Fixing Big Fat Fish's AI amnesia, saving tokens" at 257k plays.

Second, difficulty. You had to run DeepSeek Harness and the SnowLuma protocol server together, then hand-edit config.json. And people who wanted to raise fish were, in the author's own words, often "eleven- or twelve-year-old kids."

QQ-agent went after both problems: each new message opens a separate short session that gets discarded afterward, anything worth remembering long-term goes into a memory tool, and the cost per call is fixed at a small amount. It adds a "response level" slider and a cost ledger split by session and by model, so you can see where the money goes. It also no longer needs DeepSeek Harness, and any OpenAI-compatible model works. On the difficulty side, it became a graphical interface with a health-check card in the console, with no command line needed at any point.

The author was candid about the results: "That old project's prompt was what made it work well. Without that effect I'd probably have dropped it after two days." The personality half of the credit belongs to the upstream; the saving money and being easy to install were its own.

Where the stars came from: a "show off" comment in a comment section

In Sep 6, in the "QQ-agent Development Notes," the author wrote how it started: "I just saw a QQ bot project and set one up to play with… after I modified it I found it really did save a lot of cost, so I went to the comments to show it off (lol)… but too many people were watching, and if this was too half-baked it wouldn't look good, so I thought about it and added more." I couldn't find which platform or video that comment was under; from context it's most likely Bilibili.

The growth path after that was clear:

• Rode the upstream's growth together. From Oct 19 to 28, qq-bridge went from 1,207 to 3,150 stars; in the same week QQ-agent went from 550 to 979, peaking at +62 (10-21) in one day. Both curves ate the same wave of traffic; it caught the people qq-bridge's configuration and bills had scared off.
• Bilibili tutorials with cloud drive links. In Oct 24, a video called "Connecting Big Fat Fish to QQ on Mobile, a Tutorial" had a Quark cloud drive link for the QQ-agent mobile setup in its description (the video has since been deleted, per a record from one user's backup).
• Long-tail search. Articles on CSDN such as "What is QQ-agent? A complete guide to a local AI bot that mingles in groups like a real person."
• The marketplace is what keeps people around. The author says adding "feedback collection" and "quote uploads" was "kind of trying to make it something like a community." Before shutdown, the marketplace had 346 approved personas, 564 Skills, and 799 plugins (one user, 10-30, made a backup at sabercomo/qq-agent-backup).

No paid acquisition, no big-account endorsements, and no discussion I could find on X. During that stretch the author "kept sitting up all night working on it," and explained why: "I'm currently in a long stretch without work… I need something like this to make me feel I'm still 'alive in a useful way'."

GitHub, 10-09 to 10-29; after the repo was deleted, no more numbers.

Who it's competing with

Many projects bring fish into QQ, but most are plugins for DeepSeek Harness (dsh-aurabot-qq at about 440 stars, plus a string of dsh-qq-* repos with dozens to over a hundred stars), and you have to install DeepSeek Harness first. Tencent's official dsh-qqbot (546 stars) is compliant and works with a QR scan, but it only replies when @-mentioned, so it can't blend into a group. The veteran framework AstrBot (46k stars) doesn't touch the protocol layer and is doing fine, but users complain it "feels too much like a bot."

In 9 and October, a Windows desktop app that "doesn't depend on DeepSeek Harness, keeps costs under control, and is purely graphical" was almost only QQ-agent. The same-architecture qq-agent-plus was turned into a Linux server version, has 104 stars since Sep 18, and is still being updated; Tulpa (510 stars) uses MCP to hook QQ, WeChat, and Bilibili to any agent, and its README says it drew on QQ-agent. There are seven or eight more mirrors and forks.

Why it stopped: it was living inside someone else's groups

Its entire selling point rests on "logging a real person's QQ number into a protocol client and impersonating a group member." That's exactly what Tencent is going after. The author's read was that after this cleanup "the only new entrants left are QQ-Bridge and a few with very small user bases"; he also understood Tencent's side: "Some people are even reselling QQA… maybe stopping QQA from being used for scams that target those eleven- or twelve-year-old kids was the real goal." What Tencent actually did (account bans, a formal notice, or tighter risk controls) isn't backed by any official or media account, only the author's own statements.

A user in the comments offered another route: "Just switch the framework to the official bot API. AstrBot only does the agent framework and never touches the protocol, and it's fine!" The author is now building a new project called OmniAgent, the page says "under development", and it's not clear yet whether it can win back the original crowd.

If it were you

C There's something to learn here, but following it isn't worth it. Its 998 stars rode two things you can't get: the QQ buzz from the big fish (Dashboard Fish) wave, which is fading, and the teen users on QQ who wanted to raise fish, a door Tencent has since shut. Building a QQ version now will most likely get banned before you collect any stars.

Who it's for Developers who want stars for their open-source project and already have a popular upstream project being criticized for being "expensive" or "hard to install". You're meant to learn its method, not copy its topic.

Who it's not for Anyone wanting another QQ or WeChat personal-account bot (the platforms don't allow it, and the author already tried it for you). Anyone treating it as a business (the author himself thinks people reselling it or using it for gray-market purposes was one of the reasons for the crackdown).

How to do it Find a popular project whose stars are climbing and whose issues are full of complaints about cost or deployment, then build an improvement you can state in one line: "Same results, X times fewer tokens", "No command line, just double-click and go". Architecturally, QQ-agent's cheap approach of "one short session per message + a memory tool" can be carried straight over to platforms with open bot interfaces like Discord or Telegram, or you can do what AstrBot does: only build the framework and let users wire up the protocol themselves.

How to get users Same as it did: go to the upstream project's comments, issues, and tutorial videos and let measured numbers do the talking; make an illustrated tutorial for beginners; open a place early for users to upload personas and plugins, since user-made content is the best way to keep people.

When the window closes This "ride the upstream's growth" play only works during the weeks the upstream is gaining stars. The big fish's heat began to be interrupted by the crackdown starting late October, so to ride it you'd estimate launching before year's end; with any other hot upstream, you'd need to show up in the one or two weeks when it's growing fastest.

What would mean I got this wrong ① A Discord- or Telegram-based big-fish bot focused on "saves tokens, mixes in like a real person" gets over a thousand stars within a month, which would mean the platform-switch route is wider than judged here and worth upgrading; ② QQ officially opens an interface that lets bots speak without being @'d, meaning that door has reopened; ③ the author's OmniAgent quickly brings back the original personas and plugin community, which would mean the community is worth more than the platform.