← all repositories
HKUDS/Vibe-Trading

A trading agent that actually fears its own decisions

Vibe-Trading wraps LLMs in guardrails, mandates, and kill switches before letting them near real money.

11.1k stars Python Domain AppsAgentsLLMOps · Eval
Vibe-Trading
Velocity · 7d
+164
★ / day
Trend
steady
star history

What it does Vibe-Trading is a Python-based trading agent framework with a FastAPI backend and React frontend. It connects to multiple brokers (IBKR, Robinhood, Tiger, Alpaca, OKX, Binance, Futu, Longbridge) for data and—optionally—order execution. The system runs backtests, spawns multi-agent swarms for research tasks, and exposes everything through CLI, REST API, MCP, and a web UI.

The interesting bit The project treats autonomous trading as a safety engineering problem, not a feature. Live trading requires a user-committed mandate with hard bounds on symbols, order size, exposure, leverage, and daily caps. A filesystem kill switch, preemptive flatten, mandate auto-expiry, and a full audit ledger sit between the LLM and the broker. The README repeatedly notes “experimental / use at your own risk”—a refreshing honesty in a space that usually promises effortless riches.

Key highlights

  • Connector-first architecture: Broker profiles unify read-only data and paper/live trading under one selectable config; paper vs. live is a per-broker structural guard, not a toggle.
  • Research Goal runtime: Session-scoped goals persist claims, evidence, acceptance criteria, and budgets across CLI, API, MCP, SSE, and Web UI—treating agent work as auditable tasks rather than chat history.
  • Swarm with actual safety: DAG execution blocks downstream tasks on upstream failure; workers can call external MCP servers with pinned trust boundaries; stale-run reaper and heartbeat keepalive prevent phantom “running” statuses.
  • Pre-flight validation for LLM-generated code: Signal engines get AST-checked for circular imports, missing methods, and wrong return types before instantiation, with actionable JSON errors instead of tracebacks.
  • 36 MCP tools exposed for external integration, including progress heartbeats and explicit reap_stale_runs() cleanup.

Caveats

  • Longbridge connector is read-only and paper-only because its API lacks a runtime paper/live discriminator.
  • The Robinhood and newer broker connectors are marked experimental; the project is moving fast (v0.1.9 released June 2026) with frequent breaking refactors.
  • IBKR remote MCP path is currently a seeded OAuth probe awaiting stable read tool names—local TWS/IB Gateway is the solid path.

Verdict Worth a look if you’re building LLM-adjacent trading tools and need a reference for how to cage the agent before it cages your capital. Skip if you want a polished, battle-tested execution platform—this is research infrastructure with guardrails, not a finished product.

heatdrop uses Google Analytics to see which pages get read — nothing else. Your call. How we handle data.