An autonomous trading agent that perceives, decides, executes and governs its own risk.
VesperClaw closes the full agent loop on crypto markets — sensing flow, forming a thesis, paper-trading the call, and stopping itself before drawdown becomes damage. Built end-to-end for the Bitget AI Base Camp Track 1.
Autonomous perception → decision → paper execution → risk governance.
VesperClaw runs a closed control loop on every cycle. Each stage is auditable on the live dashboard, with structured outputs the next stage consumes.
Market Perception
Streams spot, funding, depth, on-chain flow and social signal into a unified state vector.
- + Price & volume aggregation
- + Funding / OI skew
- + Social & meme signal feed
Decision Engine
LLM reasoning over the state vector emits a structured intent: side, size, conviction, invalidation.
- + Thesis with rationale
- + Conviction score 0–1
- + Explicit invalidation level
Paper Execution
Trades are minted into a paper book against live mid prices — same plumbing the live router will use.
- + Mark-to-market fills
- + Slippage & fee model
- + Per-trade audit log
Risk Governance
AgentVault gates every intent; Profit Guard supervises live exposure and can flatten the book.
- + Pre-trade rule firewall
- + Drawdown circuit-breaker
- + Hard kill-switch
AgentVault. The trade either passes the rules — or it never reaches the book.
AgentVault sits between the decision engine and execution. It is a deterministic policy gate: every intent the model produces must satisfy a sequence of typed rules. Failures are logged with reasons, not silently dropped.
intent { side:"long", sym:"SOL/USDT", size:0.85, conv:0.71 } └─ check notional-cap PASS notional 1380 ≤ 2000 └─ check symbol-whitelist PASS SOL/USDT ∈ allow └─ check leverage-ceiling PASS 1.0× ≤ 2.0× └─ check conviction-floor PASS 0.71 ≥ 0.55 └─ check cooldown PASS last_trade 312s ago └─ check vol-regime WARN ATR z-score 2.1 → size×0.5 verdict ALLOW · adjusted size 0.42 · reason vol-attenuation
Halve size, widen entry filters, prefer mean-revert setups only.
Flatten directional book, hold stables, decision engine paused.
All execution disabled until human re-arm via operator cockpit.
Profit Guard turns drawdown into a stop signal — not a debate.
Profit Guard watches realized + unrealized PnL against the high-water mark. Three tiered triggers progressively dial down agency: from cautious sizing, to flat book, to full lockdown. It is the agent's own seatbelt.
- ● High-water mark tracked per session
- ● Triggers are config, not magic numbers
- ● Every state change is logged to the trade journal
PnL and a full trade log — receipts for every call.
Every intent that survives AgentVault is paper-filled and journaled. The live dashboard shows the open book, closed trades, equity curve and per-strategy attribution.
An auditable journal of every paper call.
VesperClaw keeps an auditable paper-trading log with timestamp, trading pair, direction, price, quantity, PnL, and account balance change. The live dashboard exposes the current paper book, closed trades, refusals, Profit Guard state, and prediction-market decisions.
World Cup board — the agent treats odds as a market.
VesperClaw ingests prediction-market odds as a parallel asset class, scoring mispricings against its own probability model. The World Cup board is the live demo surface.
Trending meme coins, scored before they trend on the front page.
Meme Radar fuses on-chain volume deltas, social velocity, and holder dispersion into a single Heat score. The agent uses it to size meme exposure asymmetrically — small entries, fast exits, AgentVault-capped.
Built listing-ready for the Bitget Agent Hub.
VesperClaw is packaged against the Track 1 trading-agent contract: clean intents, deterministic risk gates, structured journaling, and an operator surface a non-builder can run.
Agent manifest
Capabilities, intents and risk policy described in a single declarative file.
Structured intents
Every decision is a typed JSON object — parseable by Hub routers.
Deterministic gate
AgentVault produces audit traces the Hub can inspect per call.
Kill-switch
Profit Guard exposes a single hub-level disable endpoint.
Operator cockpit
Streamlit dashboard delivers a non-developer surface for inspection.
Trade journal
Per-trade lineage from signal → intent → fill → PnL.
Paper benchmark
Reproducible PnL track measured against BTC-buy-and-hold.
Live router stub
Same execution surface ready to be swapped from paper to live.
Open the live cockpit and watch the loop run.
No login. No keys. The agent is currently running paper trades against live markets — the dashboard updates as it perceives, decides, and acts.