What happens when you give an AI model access to your brokerage account and tell it to trade? VibeTrade, a new open-source project, is built around that exact premise.
The tool runs locally on your machine and uses Anthropic's Claude models to execute trading strategies you describe in plain English. You write something like "buy when RSI drops below 30 and sell at 5% profit," and the agent converts that into a structured Playbook with entry rules, exit criteria, position sizing, and risk guardrails.
How the Trading Loop Works
VibeTrade runs a 30-second heartbeat cycle that checks market conditions without calling the AI model on every tick. Price conditions and time-based triggers evaluate in pure JavaScript. The LLM only gets invoked when a condition fires and the agent needs to reason about whether to act.
The model stack splits work between Claude Sonnet for complex reasoning (should I enter this trade?) and Claude Haiku for lightweight condition checks. The system ships with 20 built-in tools covering live quotes, candlestick data, technical indicators, news scanning, and order book access.
The Approval Gate System
This is the part that keeps it from being a fully autonomous money-losing machine. Before any trade executes, VibeTrade presents a structured approval card showing the exact trade and every signal that led to the recommendation. Three modes are available: in-chat approval (you confirm each trade), async approval (notifications), and autonomous mode within predefined guardrails like maximum position size.
Every trade gets logged to an immutable journal with the agent's reasoning at decision time, not a post-hoc summary.
Current Limitations
Right now VibeTrade only supports Dhan, an Indian broker, with more integrations on the roadmap. You need Node.js 20+, an Anthropic API key, and Dhan credentials. All data stays local in ~/.vibetrade/ - your keys, credentials, and trade history never leave your machine.
The project is MIT-licensed and available on GitHub. A hosted version with always-on monitoring is listed as coming later.
Automated trading with LLMs is still experimental territory. Models hallucinate, markets move fast, and a confident-sounding AI recommendation is not the same as a sound one. The approval gates are a sensible design choice, but anyone running this with real money should treat it as a research tool, not a trading system they trust unsupervised.