Odds Pull
PackieY/odds-pull
Puxa odds de esportes multi-plataforma e gera relatório HTML em segundos
Ver no GitHubSkill autossuficiente que converte uma frase em linguagem natural em um relatório HTML de odds de esportes cross-platform. Integra dados de Polymarket, DraftKings, FanDuel e Kalshi, aplicando filtros inteligentes e detecção de anomalias para insights financeiros imediatos.
Instalação
npx convez-ai PackieY/odds-pull1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Análise rápida de odds para apostas desportivas
- Detecção de anomalias e disparidades entre plataformas
- Geração automatizada de relatórios de mercados de previsão
Veredito

Solução robusta e bem arquitetada para traders e analistas que precisam monitorar odds multi-plataforma programaticamente. Exige configuração local (API keys, Python) mas oferece pipeline completo de coleta, matching inteligente e relatórios visuais prontos para ação.
README
A tradução para o português deste README ainda está na fila.
odds-pull skill
A portable, self-contained Claude Code skill that turns one natural-language sentence into a multi-platform sports-odds HTML report.
What it does
You say:
帮我拉一下 NBA 这周末的 odds 数据 compare odds for tonight's NBA games pull mainstream odds for the next 3 days
…and your AI:
- Parses the sport + time window from your sentence
- Fetches odds from Polymarket, DraftKings, FanDuel, Kalshi
- Cross-platform matches games (with LLM fallback for tricky cases)
- Filters to main-event games (≥ 3 platforms, volume ≥ $10K)
- Generates a self-contained HTML report — opens in your browser
- Summarizes the key signals (vig hotspots, anomalies) in chat
Compatibility
| Environment | Status |
|---|---|
| Claude Code (CLI or VS Code extension) | ✅ Auto-triggers; native sub-agent flow |
| Cursor / Aider / Continue / Codex CLI / WorkBuddy / etc. | ✅ Works — point your AI at SKILL.md, it'll follow the steps. The two LLM sub-agent steps (3 + 3.5) have explicit "if you have no Agent tool, inline-execute the prompt" fallback notes. |
| Pure Python invocation (no AI orchestrator) | ✅ for data + HTML pipeline (pull_odds.py → html_writer.py); ⚠️ skip the LLM merge / financial-analysis steps unless you wire your own LLM glue |
| LLM with no shell / file IO (browser-only chat) | ❌ Won't work — fetching needs local Python execution |
Smoke test (post-install)
bash scripts/smoke_test.sh
Verifies Python deps + .env + a small NBA-only fetch + HTML render in ~30 seconds, without needing the LLM agents. Use this after a fresh clone or after editing vendored code.
Quick install
See INSTALL.md for full instructions. The TL;DR:
# Manual install
git clone <REPO_URL> ~/.claude/skills/odds-pull
cd ~/.claude/skills/odds-pull
pip install -r requirements.txt
echo "ODDS_API_KEY=your_key_from_the-odds-api.com" > .env
chmod 600 .env
bash scripts/smoke_test.sh # ← verify before first real use
Or paste the AI-assisted install prompt from INSTALL.md §2 into your AI.
Architecture
User: "帮我拉一下 NBA 这周末的 odds"
│
▼
Claude (orchestrator) reads SKILL.md → 5-step recipe
│
├─ Step 1 [LLM] Parse intent → CLI args
│ ("这周末" → --hours-max 96)
│
├─ Step 2 [Bash] scripts/pull_odds.py
│ ├─ load_dotenv() — get ODDS_API_KEY
│ ├─ fetch_all_markets() — parallel pull
│ │ PM (~15s) + The Odds API (~2s) + Kalshi (~2s)
│ ├─ build_game_rows() — cross-platform join
│ ├─ Apply --min-volume-usd + --min-platforms filters
│ └─ Dump JSON to /tmp/odds_pull.json
│
├─ Step 3 [Agent] If orphans non-empty:
│ Spawn general-purpose sub-agent with orphan list
│ → returns JSON pairs (with confidence)
│ → write back to JSON's llm_matches field
│
├─ Step 4 [Bash] scripts/html_writer.py
│ ├─ Apply LLM-merged orphans
│ ├─ Detect anomalies (PM-missing-Spread, line mismatch,
│ │ high vig, low-confidence LLM match)
│ ├─ Compute vig heatmap (matched-only)
│ └─ Render self-contained HTML → output path
│
└─ Step 5 [Text] Tell user file path + 3 key anomalies + 1-line insight
Default behaviour (per design spec)
--sports mainstream— NBA + NFL + MLB + NHL + SOCCER (off-season leagues return empty; that's fine)--hours-min 0 --hours-max 72— 3-day window--min-platforms 3— only show games seen on ≥ 3 platforms (i.e. real cross-platform comparisons; drops single-platform long-tail markets)--min-volume-usd 10000— drop PM/Kalshi markets with < $10K lifetime volume (thin liquidity skews vig calculations)
Override via natural language:
| User says | Becomes |
|---|---|
| "看看所有运动" | --sports all |
| "包括 PM 那种小盘子" | --min-platforms 1 --min-volume-usd 0 |
| "只看 NBA 今晚" | --sports nba --hours-max 12 |
Files
| Path | Purpose |
|---|---|
SKILL.md | Trigger keywords + 5-step orchestration recipe |
INSTALL.md | Full install guide (manual + AI-assisted) |
README.md | This file (architecture overview) |
requirements.txt | 3 Python deps (aiohttp, Brotli, python-dotenv) |
.env | YOU create this; holds ODDS_API_KEY (not committed) |
scripts/pull_odds.py | Fetch + cross-platform join → JSON |
scripts/html_writer.py | JSON → self-contained HTML report |
lib/oddspull_vendored/ | Vendored core libraries (~5900 lines) |
Why vendored core libraries?
The skill ships with a frozen copy of the data-layer code from the parent
odds_comparison project (under lib/oddspull_vendored/). This is
deliberate — it makes the skill fully self-contained:
- Your colleague clones this repo to
~/.claude/skills/odds-pull/and it works immediately. No need to also cloneodds_comparison. - The vendored code carries 100+ months of bug fixes — Kalshi fee continuous formula, strike-by-volume picker, commence_time correction, sport-aware team-name disambiguation, retry-on-5xx-with-cache-skip, etc.
- The trade-off: when the parent project fixes a new edge case (e.g. a new Kalshi API field), you'll need to re-vendor (run a sync script — TBD). But for a stable consumer, this is rare.
What's NOT in this skill
- No Excel writer — HTML is the only output format. Open in any browser. If you want Excel, copy-paste from the HTML table or open the JSON in pandas.
- No live web UI — the HTML is a static snapshot of the data at fetch
time. For continuous monitoring, see the parent project's
web/app. - No Kalshi browser-mode auth — Kalshi's public REST API is enough for the data this skill needs. No Playwright dependency.
- No PM mobile-app spread/total scraping — Polymarket's mobile sportsbook uses a third-party odds feed not exposed in their public API. Markets missing on PM but present elsewhere are flagged in the Anomalies section.
Future (Stage 2 — not yet implemented)
- Pre-flight common-sense agent: spawned at Step 1.5; uses WebSearch to check what's in season + any major news (injuries, trades, league decisions) that could affect the games being pulled
- Post-report financial-analysis agent: spawned at Step 4.5; reads the vig heatmap + anomalies and writes investment-style commentary into a new section of the HTML
To enable Stage 2 later, edit SKILL.md to add the two extra spawn steps.
No script changes needed.