A Trading Agent Framework That Actually Fears Margin Calls

What it does

Vibe-Trading is a Python platform that connects LLM agents to actual brokerage APIs—IBKR, Robinhood, Binance, OKX, Alpaca, and others—for research, backtesting, and (optionally) live order placement. It exposes a unified interface across a React web UI, an interactive Rich CLI, REST endpoints, and MCP. The system centers on persistent Research Goals that track claims, evidence, and acceptance criteria, while a swarm layer dispatches multi-agent DAGs to run tasks like backtests or factor analysis.

The interesting bit

The project treats risk containment as a core engineering problem rather than a disclaimer: live connectors require user-committed mandates with symbol universes, order-size caps, and daily limits, plus a filesystem-level kill switch and a fail-closed pre-trade gate. Even the swarm layer can invoke external MCP servers, but the trust boundary is explicitly pinned so workers don’t wander off-script.

Key highlights

• Connector-first architecture supports paper and bounded-live trading across IBKR, Robinhood, Alpaca, Binance, OKX, Futu, Tiger, and Longbridge (the latter is read-only/paper-only).
• LLM-generated signal engines undergo pre-flight AST and interface validation before instantiation to catch circular imports, missing methods, or wrong return types.
• Research Goals persist across sessions with evidence rows, budgets, and completion policies, accessible from CLI, REST, MCP, and the web UI.
• Swarm execution uses a DAG that blocks downstream tasks on upstream failure and includes a strict alpha benchmark with random controls to weed out beta-tracking factors.
• Session storage uses flush + fsync on every append and skips corrupted JSONL lines on read, which is the kind of paranoia you want when an AI is managing trade state.

Caveats

• Live trading is explicitly labeled “experimental / use at your own risk” for every supported broker, including the new connector batch.
• The README reads more like a high-velocity changelog than an architectural overview, so understanding the current surface requires piecing together dated news entries.
• Longbridge is currently read-only and paper-only because its API lacks a runtime paper/live discriminator.

Verdict

Worth exploring if you’re building LLM-driven quant tools and want a structured, safety-conscious bridge to real brokers. Avoid if you need a simple, proven execution layer without the agentic orchestration overhead.