A semi-automated trading system for Polymarket prediction markets. Human + AI hybrid: human places trades, AI (via Claude Research) does analysis, bot handles tier-based take-profit and stop-loss.
⚠️ Disclaimer: This is research-grade software. Trading prediction markets involves risk of loss. Use at your own risk. The author is not liable for any financial losses.
Unlike fully-automated trading bots, this system splits the work:
| Component | Role | Why |
|---|---|---|
| Scanner | Pulls markets from Polymarket Gamma API, applies hard filters (price/volume/depth/spread), generates Markdown report | Mechanical filtering of liquidity & dates |
| Human | Reviews the scan report, runs Claude Research analysis, decides whether to enter | AI provides analysis, human makes final judgment |
| Trader (you) | Manually places limit orders on Polymarket website (as maker, saves taker fee) | Avoids stupid mistakes, captures rebates |
| Monitor (bot) | Every 3 minutes checks positions, executes T1/T2/T3 take-profit and W1/W2/W3 reverse-stop-loss tiers | Mechanical execution of pre-defined rules |
| Re-evaluation | When a position hits 80% of TP target, dashboard shows yellow alert prompting human to re-run Claude Research | Avoids exiting too early on momentum-driven events |
~/polymarket-semi-auto/ ├── main.py # Entry: starts monitor + Flask ├── .env # Credentials (DO NOT commit) ├── semi_auto.db # SQLite (events / tier_sold / position_meta) ├── bot.log # Runtime log └── modules/ ├── dashboard.py # Flask web UI + API routes ├── scanner.py # Market scanner + report generator ├── monitor.py # Position monitor + tier rules ├── executor.py # CLOB client wrapper (sell with FAK + best_bid) ├── db.py # SQLite interface ├── prompts.py # Claude Research prompt templates └── tags.py # 22 whitelisted tags + blacklist
- Two scan modes: keyword search or tag-based discovery (22 whitelisted Polymarket tags grouped by Tier 1/2/3)
- Three filter ranges (standard / medium / wide) with progressively relaxed liquidity thresholds
- Hard filters: price range, volume, order book depth, spread, settlement window, ambiguous resolution words
- Dual blacklist (tag-based + keyword-based with word-boundary matching)
- Reports grouped by event for cluster analysis
- Take-profit tiers: T1/T2/T3 at +5pp/+10pp/+15pp from entry, selling 25%/35%/40% respectively
- Reverse stop-loss tiers: W1/W2/W3 with gap-relative dynamic distances
- Time stop: ≤2 days to settlement + price barely moved → full exit
- Small-edge mode: when gap < 10pp, single exit at mp + min(gap-2, 8pp)
- Direction-unified: both YES and NO positions follow "holding token's price up = profit"
- Uses FAK (Fill-And-Kill) order type — fills immediately or cancels, never rests
- Pulls best_bid as limit price for slippage protection
- Strictly validates
result.successANDmakingAmount > 0(avoids "fake success" bug) - Logs partial fills as warnings
- When position hits 80% of TP target, dashboard shows yellow re-eval badge
- One-click prompt copy → paste to Claude Research → choose A (uplift TP) / B (skip) / C (exit early)
- Bot does NOT auto-uplift; human decides based on Claude's findings
- One re-evaluation per position lifetime (status machine: pending → done_*)
- Tab UI: keyword scan / tag scan
- 22 tag chips colored by Tier (green/blue/purple)
- Position table with TP input + manual save
- Trigger price display: T1/T2/T3 (green) and W1/W2/W3 (red)
- Partial refresh every 30s (no full page reload)
- Real-time log tail
- Python 3.12+
- Polygon wallet with USDC funded on Polymarket
- Polymarket account using EOA, Magic/Email, or Proxy Wallet
git clone https://github.com/RobinVico/polymarket-semi-auto.git
cd polymarket-semi-auto
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txtcp .env.example .env
# Edit .env with your credentials| Variable | Required | Description |
|---|---|---|
POLY_PRIVATE_KEY |
Yes | Wallet private key (0x...) |
POLY_FUNDER |
Yes | Funder address (0x...) |
POLY_SIGNATURE_TYPE |
Yes | 0=EOA, 1=Magic, 2=Proxy |
POLY_API_KEY/SECRET/PASSPHRASE |
No | Auto-generated on first run |
# Foreground
python3 main.py
# Background
nohup python3 main.py > output.log 2>&1 &Open http://localhost:5050 in your browser.
- Scan — pick a tag or enter keyword, choose range (standard/medium/wide)
- Copy prompt to Claude — click "Copy for Claude" button
- Run Claude Research — paste prompt, wait 5-15 min for analysis
- Review recommendation — Claude outputs human-readable cards (slug / direction / price / raw estimate / calibrated estimate / edge / confidence)
- Place order manually — go to Polymarket website, place Limit order as maker
- Save TP — back in dashboard, fill TP value (calibrated probability) for that position
- Bot takes over — monitors price every 3 minutes, executes T/W tiers automatically
Tier 1 (11 tags) — strongest analyst edge Iran, Israel, Ukraine, Ukraine Peace Deal, Russia, China, Taiwan, Geopolitics, Middle East, World, Foreign Policy
Tier 2 (10 tags) — moderate edge Trump, Trump Presidency, SCOTUS, Politics, US Politics, AI, OpenAI, Tech, Science, Venezuela
Tier 3 (1 tag) — reverse-bet (sell NO preferred) Awards
Blacklist — auto-rejected: Crypto/Bitcoin/Ethereum/Solana/Sports/Soccer/NFL/NBA/Hurricane/Weather/Fed/CPI/Inflation/...
| Item | Standard | Medium | Wide |
|---|---|---|---|
| Price range | 5%-92% | 4%-93% | 3%-95% |
| 24h volume | ≥$500 | — | — |
| Total volume | — | ≥$2000 | ≥$3500 |
| Depth | ≥$3 | ≥$2.5 | ≥$1.5 |
| Spread | ≤5pp | ≤6pp | ≤8pp |
| Settlement | 5-30d | 4-45d | 3-75d |
| Ambiguous words | strict | strict | strict |
- Small edge (gap < 10pp): single exit at
mp + min(gap-2, 8pp) - Tiered (gap ≥ 10pp):
- T1 →
mp + 5ppsell 25% - T2 →
mp + 10ppsell 35% - T3 →
min(mp + 15pp, tp - 2pp)sell remaining 40%
- T1 →
- W1 →
max(5pp, min(gap×0.33, 8pp))sell 30% - W2 →
max(10pp, min(gap×0.67, 14pp))sell 40% - W3 →
max(15pp, min(gap×1.0, 20pp))full exit
Distance to settlement ≤ 2 days AND |cur - mp| < 5pp → full exit
| Route | Method | Purpose |
|---|---|---|
/ |
GET | Dashboard |
/api/control |
POST | scan / scan_tag / check / refresh |
/api/snapshot |
GET | Position snapshot for partial refresh |
/api/scan_report |
GET | Latest scan markdown |
/api/full_prompt |
GET | DISCOVERY_PROMPT + scan report joined |
/api/record_position |
POST | Save TP with direction sanity check |
/api/update_tp |
POST | Update TP |
/api/reeval_prompt |
GET | Generate re-eval prompt for a position |
/api/mark_reeval |
POST | Mark uplift / skip / close |
/api/logs |
GET | Last 80 lines of bot.log |
- Python 3.12 + Flask + SQLite
- py-clob-client for Polymarket CLOB
- Polymarket Gamma API for market discovery
- Claude Research (via copy-paste, no API integration) for market analysis
- Manual order entry: human must place orders on Polymarket website (no auto-buy)
- Claude Research subscription required for analysis (Claude Max plan or equivalent)
- VPN may be required depending on jurisdiction
- Maker rebates only on Finance category in Polymarket — geopolitical/political markets pay full taker fee
- No portfolio sizing — designed for fixed-size manual entries (Kelly criterion not implemented)
- Full Technical Documentation (Markdown) - System architecture, design rationale, all SL/TP rules, prompt design, bug history
- Technical Documentation (Word) - Same content in formatted Word document
MIT — see LICENSE
This software is provided as-is. Prediction markets carry financial risk. The author makes no warranty about profitability, accuracy of analysis, or correctness of execution. Test thoroughly with small amounts before scaling.
