A Discord bot that discovers open-source GitHub repositories, evaluates them for latent business and product potential using an LLM venture analyst, and streams approved findings directly to your Discord channel.
06:00 UTC ──► Daily keyword scan (run)
├─ 10 keywords from deterministic weekly cycle
├─ Per keyword: fetch top repos, take first 3 passing pre-filter
├─ is_seen(url) check — skip if already in DB
├─ LLM venture-analyst eval (README + structured prompt)
├─ KEEP + score ≥ 7.5 → upsert to DB + post to Discord
└─ Posts up to 3 KEEPs, ~30 LLM calls max
10:00 UTC ──► Trending scan (run_quick, mode=trending)
├─ GitHub: stars≥10 created in last 7 days (2 queries)
├─ Evaluate one by one until 2 KEEPs posted
└─ Max 20 LLM evals (quick_max_eval cap)
14:00 UTC ──► Popular scan (run_quick, mode=popular)
├─ GitHub: stars≥100 pushed in last 3 days
├─ Evaluate one by one until 2 KEEPs posted
└─ Max 20 LLM evals (quick_max_eval cap)
/scan [keyword] ──► Quick scan (run_quick, mode=keyword)
├─ Keyword or 3 random from config
└─ Stop on first KEEP ≥ 7.5, max 1 post
/trending, /popular ──► Same as cron runs, triggered on demand
DB model: only KEEPs are stored. REJECTs are never written. is_seen(url) prevents re-evaluating any repo already in DB.
- Python 3.12+
- A Supabase project
- A Discord bot application (create one here)
- A GitHub personal access token
- A Google AI Studio API key (Gemini)
- Go to Discord Developer Portal
- Click New Application, give it a name
- Go to Bot tab → click Reset Token → copy the token
- You do not need Message Content Intent (slash commands only)
- Go to OAuth2 → URL Generator:
- Scopes:
bot,applications.commands - Bot Permissions:
Send Messages,Embed Links,Read Message History
- Scopes:
- Copy the generated URL and open it to invite the bot to your server
- Right-click the target channel → Copy Channel ID (enable Developer Mode in Discord settings if needed)
Run the SQL in supabase/schemas/repos.sql against your Supabase project (via the SQL Editor in the dashboard).
To clear the database for a fresh start:
truncate table public.repos restart identity;cp .env.example .env| Variable | Required | Description |
|---|---|---|
DISCORD_TOKEN |
✓ | Your Discord bot token |
DISCORD_CHANNEL_ID |
✓ | Channel ID where the bot posts |
GITHUB_TOKEN |
✓ | GitHub personal access token |
SUPABASE_URL |
✓ | Your Supabase project URL |
SUPABASE_KEY |
✓ | Supabase service role key (not anon) |
GEMINI_API_KEY |
✓ | Google AI Studio API key |
LLM |
✓ | Model name (e.g. gemini-2.5-flash) |
All other settings have defaults in bot/config.py:
| Setting | Default | Description |
|---|---|---|
REPOS_PER_KEYWORD |
3 |
Repos evaluated per keyword in daily scan |
MAX_POST_PER_RUN |
3 |
Max KEEPs posted per daily keyword scan |
QUICK_MAX_EVAL |
20 |
Max LLM evals per trending/popular run |
MIN_SCORE_TO_POST |
7.5 |
Minimum score to post to Discord |
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python -m bot.main| Command | Cooldown | Description |
|---|---|---|
/scan [keyword] |
30s | Quick scan — evaluates one by one, posts first KEEP ≥ 7.5 |
/trending |
60s | Scan recently created repos with fast star growth (last 7 days), up to 2 posts |
/popular |
60s | Scan established repos (≥100 stars) with active recent pushes, up to 2 posts |
/crontest |
600s | Manually trigger today's 10-keyword daily cycle |
/keywords |
— | List all 70 configured keywords and the daily cycle logic |
/top [count] |
— | Show the top N highest-scoring repos from the database |
/repo <url> |
30s | Analyze a specific GitHub repo on demand |
Uses a deterministic weekly cycle over 70 keywords:
- Keywords are shuffled once per week (seeded by ISO week number)
- Today's weekday selects a slice of 10 from that shuffled list
- Full coverage across all 70 keywords over 7 days, no repetition within a week
Per keyword: one GitHub API call for repos created in the last 14 days with ≥5 stars, sorted by stars. Takes the first 3 repos that pass the pre-filter, evaluates each. Only KEEPs scoring ≥ 7.5 are stored and posted, up to 3 per run (~30 LLM calls max).
Broad signal scans not tied to keywords:
- Trending: repos created in the last 7 days with ≥10 stars (two threshold queries)
- Popular: repos with ≥100 stars pushed in the last 3 days
Both evaluate candidates one by one until 2 KEEPs are posted or the 20-eval cap is hit.
The venture analyst prompt has five steps: understand → filter → identify latent opportunity → productize → score.
Rejects on any single disqualifier: purely academic, dataset/weights only, no identifiable paying user, personal/homework project, or abandoned. A repo must have a credible — not just conceivable — path to revenue.
Scores are weighted: (0.5 × business_potential) + (0.3 × novelty) + (0.2 × ease_of_mvp). Only repos scoring ≥ 7.5 are posted to Discord.
Only KEEPs are written to the DB. is_seen(url) does a single DB read before any LLM call — if the URL exists, the repo is skipped immediately at zero cost. REJECTs are never stored, so the DB stays small and meaningful.
Using gemini-2.5-flash with billing enabled:
| Run | Max evals/day | Est. cost/day |
|---|---|---|
| Keyword scan (06:00) | 30 | ~$0.01 |
| Trending (10:00) | 20 | ~$0.007 |
| Popular (14:00) | 20 | ~$0.007 |
| Total | 70 |
bot/
├── main.py # Entry point: Discord bot + APScheduler (3 cron jobs)
├── config.py # Settings (env vars + defaults)
├── cogs/
│ └── commands.py # /scan, /trending, /popular, /crontest, /keywords, /top, /repo
├── modules/
│ ├── ingestion.py # GitHub API: keyword search, trending, popular strategies
│ ├── prefilter.py # Deterministic filter (stars, fork, description, age, size)
│ ├── dedup.py # is_seen(), upsert_keep(), mark_posted(), get_top_repos()
│ ├── evaluator.py # LLM venture analyst prompt + JSON parsing + retry logic
│ ├── ranker.py # Weighted score computation
│ ├── discord_poster.py # Discord embed builder
│ └── pipeline.py # run() daily scan, run_quick() triggered scans, evaluate_single()
└── db/
└── client.py # Supabase singleton client