A cognitive firewall between impulse and execution.
Adversarial AI layer that forces traders to defend every on-chain trade — with live OKX evidence — before it signs.
Built for the XerpaAI × OKX "Build X-Agent" Hackathon — May 2026.
Crypto trading interfaces are optimized for speed of execution. The market isn't the enemy. Impulse is.
74–89% of retail traders lose money. Pre-trade tools — bots, scanners, checklists, journals — fail at the same point: the moment of intent. Nothing in crypto creates forced cognition before execution.
That's the gap. Repps × OKX fills it.
A four-round adversarial rehearsal over Telegram, before any trade signs:
| Round | Forces the user to articulate... |
|---|---|
| 1 · Thesis | Why are you buying this? (in one sentence, before evidence is shown) |
| 2 · Defend | Live OKX evidence cited as the attack. Dev wallet history, smart money flow, holder concentration. Defend this. |
| 3 · Exit | Three concrete numbers: take-profit, stop-loss, max hold duration. "See how it goes" is a fail. |
| 4 · Size | Position as % of portfolio, then type I can lose this verbatim. Friction is the feature. |
The state machine cannot be shortcut. Each round captures the user's reasoning; the verdict at the end is informed by the cumulative pattern.
Three states. No fake-precision numeric scoring.
| 🟢 | Go | Execute via OKX agentic wallet (simulated in v0.1) |
| 🟡 | Pause | 60-minute cooling-off. Revisit when reasoning is sharper. |
| 🔴 | Abort | 24-hour mandatory cooling-off. Hard reset. |
The execute button only surfaces on Go. Pause and Abort persist in user state — the same token cannot be re-rehearsed during the cooldown window.
The moat is execution-context awareness:
- Live on-chain evidence pulled at the moment of intent, not from training data
- Forced protocol — the four rounds aren't a suggestion. The state machine can't be shortcut.
- Asset-class differentiation — meme tokens trigger rug-pattern protocols, mid-caps focus on tokenomics, majors focus on allocation discipline
- Behavioral memory (v0.2) — every rehearsal compounds. Override rate, sizing variance, hype-flag frequency become user-specific patterns.
- Execution gating — ChatGPT can give advice; it can't sign a transaction. Repps gates execution behind the rehearsal.
Repps × OKX integrates with OKX OnchainOS via REST APIs.
| Data | Status | Source |
|---|---|---|
| Token search & metadata | ✅ Live | dex/aggregator/all-tokens |
| Live price (latest candle close) | ✅ Live | dex/market/candles |
| Security scan & contract risk | ✅ Live | security/token-scan |
| Holder concentration | ⚪ Illustrative | OnchainOS CLI-only — REST exposure pending |
| Smart money flow | ⚪ Illustrative | OnchainOS CLI-only — REST exposure pending |
| Dev wallet history | ⚪ Illustrative | OnchainOS CLI-only — REST exposure pending |
The three illustrative data points are exposed only via OKX's CLI-based Skills runtime, not the REST surface that a standalone Node process can consume. The integration code is built against the documented skill contracts and ready for live data when REST exposure lands. See ROADMAP.md for v0.2+ integration plans.
- Open Telegram and search for @reppscrypto_bot (or tap the badge at the top)
- Send
/rehearse BONK 500 SOL— or any Solana token and size you're considering - Defend the trade through four rounds
- Get a Go / Pause / Abort verdict
- Cleared? Execute. Paused? Sit with it. Aborted? Move on.
Available commands:
/rehearse <token> <size> <unit> Start a rehearsal
/journal View past rehearsals
/abort Cancel current rehearsal
/help Show this list
v0.1 supports Solana mainnet with SOL-denominated sizing. Multi-chain (ETH, Base, BSC, Polygon, Arbitrum) and USD-denominated sizing arrive in v0.2.
- Bot: TypeScript + grammY (Telegram framework)
- Rehearsal logic: Claude Sonnet 4 via the Anthropic API with JSON-mode verdict parsing
- On-chain data: OKX OnchainOS REST API + HMAC-SHA256 auth (production wallet-auth keys)
- State persistence: Supabase (Postgres + service-role auth)
- Hosting: Node process, deployable to Railway / Render / any Node-friendly platform
Repps × OKX is part of the Repps AI family — cognitive rehearsal for consequential decisions.
The visual register is mission-control dark UI. Deep Graphite (#111315) background, Bone White text, Electric Amber for evidence callouts and accents. Typography: JetBrains Mono for the wordmark and numbers, Inter for prose. The aesthetic sits between Linear, Arc Browser, and Palantir — warmer than Linear, more product than Palantir.
The voice is adversarial, evidence-first, calm under pressure. Never preachy. Never coaching. The bot states facts and asks for defenses; it does not coach, motivate, or congratulate. "Defend this." not "You should consider..."
See ROADMAP.md for the full plan:
- v0.2 — Multi-chain expansion (ETH, Base, BSC, Polygon, Arbitrum + USD-denominated sizing)
- v0.3 — Cross-chain & bridges (OKX cross-chain swap quotes)
- v0.4 — Multi-DEX comparison (Jupiter, Uniswap, Aerodrome)
- Beyond — OnchainOS deep integration, behavioral memory, Repps AI family cross-vertical pattern detection
To set expectations honestly:
- Not a trading bot. Repps × OKX doesn't pick trades, generate signals, or recommend buys.
- Not a coach. The voice is adversarial, not supportive. No "great rehearsal!" outro.
- Not a journal. Past rehearsals inform pattern recognition in v0.2+, but the product is forward-facing — what you're about to do, not what you did.
- Not financial advice. The bot challenges thinking; it doesn't direct outcomes.
| Resource | Link |
|---|---|
| Live bot | t.me/reppscrypto_bot |
| Landing page | reppsai.com |
| Repo | github.com/drock8/repps |
| Hackathon | XerpaAI × OKX "Build X-Agent" — May 2026 |
Built with:
- OKX OnchainOS — the on-chain data layer
- Anthropic Claude Sonnet 4 — the rehearsal brain
- grammY — Telegram framework
- Supabase — persistence
Part of the Repps AI family — cognitive rehearsal for consequential decisions.
Built by Derrick Siu.