Case · An open-source project handled a 300k Discord community with Chatwoot + Dify
A 18k-star OSS project went from drowning in Discord to clean AI + volunteer mod tiers handling 300k community asks.
Background#
- Project: an open-source developer tool (name redacted)
- GitHub stars: 18,000+
- Discord members: ~310,000
- Monthly new GitHub issues: ~600
- Monthly Discord questions: ~12,000
- Team: 3 full-time maintainers + 12 volunteer mods
Pre-AI pain#
The success penalty of OSS projects:
- Discord scrolls every minute — English users ask, then Chinese, Russian, Portuguese
- 80% are repeat questions — “how do I install,” “configure X,” “why does Y error” — 100× per day
- Maintainer burnout — 3 maintainers full-time answering, no code time
- GitHub issues stuck at 600+ open — many aren’t bugs, just questions
Architecture#
Discord + Chatwoot is unconventional. Here’s the wiring:
Interesting design choices#
1. AI answers ship with voting#
Dify Workflow auto-adds reactions: 👍 / 👎 / 🙋 (request human).
- 👍 ≥ 3 → sync to GitHub Discussions (knowledge accrues)
- 👎 ≥ 2 → trigger improvement loop, prompt engineer reviews that week
- 🙋 ≥ 1 → page a moderator
2. KB built by the community#
Every doc PR triggers n8n webhook → Dify KB sync. The last 12 months of GitHub Discussions answers with ≥ 5 votes are vectorized in too.
The KB isn’t maintained by core team — it’s a side-effect of community behavior.
3. Multilingual with one source#
Maintainers write English docs only. Dify prompt:
You are the community assistant for this project.
Detect user language and reply in it (zh / en / ru / pt / ja / ko / ...).
Query the KB in English.
Keep code examples as-is, don't translate.4. Time-zone-aware#
Maintainers are spread UTC-5 / UTC+0 / UTC+8. There used to be ~3-hour gaps with no one around. AI now runs 24/7; maintainers only watch the “escalated” channel.
5-month numbers#
| Metric | Before | After |
|---|---|---|
| Daily Discord questions | ~400 | ~600 (community grew) |
| Maintainer hours / day on support | 6-8 | 1-2 |
| Avg response | 47 min | 4.2 s |
| AI deflection | — | 73% |
| Mod escalation | — | 19% |
| Maintainer escalation | — | 8% (real bugs / complex) |
| GitHub open issues | 624 | 287 (many resolved on Discord first) |
| Monthly maintainer PRs | 18 | 27 (more code time) |
Biggest win: maintainer weekly coding time went from 12 hours to 35 hours. Project velocity doubled.
5 unexpected findings#
1. “Unmanned AI” attracts users#
A user tweeted “@project answered in 4 seconds, faster than my company’s internal support.” OSS support quality became a selling point.
2. AI answer quality stress-tests docs#
Where AI gets 👎 a lot, docs are unclear. That’s now the maintainers’ “fix the docs” signal.
3. Volunteer engagement rose#
Volunteers previously felt “all repeat questions, no fun.” Now they only see complex cases → more interesting. Attrition dropped from 30%/yr to 5%/yr.
4. Context windows#
Discord long threads often hit 50+ turns. Dify default 5-10 turns isn’t enough. Switched to sliding window + summarization: compress every 10 turns into a summary.
5. Emoji replies got massively positive responses#
Community users react strongly to emoji in AI replies. An “unskilled” AI is better received — ”🤖 I’m trying to answer, please verify” gets 40% more 👍 than “This is the answer:”.
Cost#
| Item | Monthly |
|---|---|
| 1× 4C/8G VPS (Chatwoot + Dify) | $50 |
| LLM tokens (DeepSeek + Qwen 14B for reranking) | $80 |
| Vector store (pgvector, free) | $0 |
| Discord Bot host (Cloudflare Workers) | $5 |
| Domain | $1 |
| Total | ~$136 / mo |
Fully open source + self-hosted + Discord-primary = extremely low hosting cost.
Advice for OSS maintainers#
- Discord + AI is a killer combo for OSS — don’t over-engineer with Stack Overflow / Discourse
- Let the community self-tune the KB — good answers auto-flow in
- AI being wrong is fine — let users 👍/👎; cheaper than preventing errors
- Wire GitHub PRs / Discussions in too — same n8n orchestration
- 3 core + AI > 30 volunteers — free core people for creative work
Don’t#
- Connect Discord Bot to OpenAI directly — bills will explode at scale
- Let AI promise commercial deployment / SLAs — must escalate
- Index all Discord history into KB — introduces stale information