Live runtime
- Agent orchestration & tool routing
- Web UI, Telegram, HTTP, MCP
- Exchange & broker integrations
- File-driven config & sessions
ClawQuant
AI-native quant lab
A self-evaluating trading agent that learns from every executed trade.
Replay. Attribution. Gated policy promotion. Quant-first. File-driven. Always learning.
Not just trading. Evolving.
Learning loop
Every closed trade feeds a closed loop: observe, attribute, benchmark, and promote only what survives replay.
01
Observe
Capture market context, features, and policy version at decision time.
02
Execute & log
Orders, fills, risk envelope — immutable record for later labeling.
03
Label & attribute
Delayed outcomes, horizon returns, entry / exit / regime / sizing fit.
04
Replay & benchmark
Historical slices, walk-forward framing — candidate vs baseline.
05
Promote
Gated policy promotion — only if better and safer.
Architecture
TS runs the system. Python learns. Rust accelerates. One product story, three execution surfaces.
Research brain
- Backtests & benchmarks
- Trade labeling & attribution
- Experiments & reports
- Self-evaluation pipelines
Speed core
- Technical indicator engine
- Causal candle replay
- Room for tick / portfolio kernels
- Performance where it counts
TS runs the system. Python learns. Rust accelerates.
Features
Serious infrastructure for builders who want AI in the loop without leaving quant discipline behind.
Dual AI provider
Claude Code CLI or Vercel AI SDK — switch at runtime without restarting the desk.
Crypto & equities
CCXT and Alpaca paths with git-like wallet workflows for staged execution.
Market intelligence
OpenBB-backed layers, analysis kit, and room for Rust-accelerated indicators.
Event log & cron
Durable JSONL events, scheduled jobs, heartbeat — operations that survive restarts.
File-driven everything
Markdown persona, JSON config, JSONL sessions — humans and agents read the same truth.
Evolution mode
Sandboxed brain vs full project access — explicit permission for self-modifying agents.
Repo structure
Multi-language layout you can deploy, fork, and extend — not a monolith masquerading as research.
# ClawQuant — high level src/ # TypeScript agent runtime ui/ # Web UI python/claw_quant/ # Research, labeling, benchmarks rust/ # indicator-engine, replay-core packages/schemas-ts/ # Shared artifacts (e.g. trade learning records) data/ # Config, sessions, future learning/* docs/ # Architecture notes
“The edge isn’t a single model — it’s the loop: replay what you did, attribute what worked, and promote only what proves out under scrutiny.”
Vision — ClawQuant
Ship the loop.
Star the repo, run the agent, and wire your first trade learning record. Elite quant lab meets AI-native execution.