VibeTrade, an open-source project published at github.com/vibetrade-ai/vibe-trade, offers a locally-run AI trading agent built around Anthropic's Claude Sonnet and Haiku models. Rather than wrapping Claude in a simple chatbot interface, the project frames itself as an infrastructure layer — a "harness" providing the six primitives an LLM needs to operate responsibly in financial markets: Market Tooling (20 built-in broker and data tools), Heartbeat (a 30-second polling loop), Triggers (condition evaluation), Permissions (a human-approval gate), Playbooks (persistent plain-English strategy documents), and Learnings (an immutable decision journal). The system currently connects to the Dhan broker for Indian equity markets, with support for additional brokers listed on the roadmap.
The architectural core of VibeTrade is its separation between cheap monitoring and expensive reasoning. The Heartbeat loop runs every 30 seconds in pure JavaScript, evaluating price and time conditions without invoking the LLM. Claude is only called when a trigger actually fires — at which point it draws on live quotes, candles, indicators, and news feeds via the Market Tooling suite before producing a recommendation. This keeps ongoing operating costs low while letting Claude weigh in only when conditions actually trigger a decision. Claude Haiku handles lighter fuzzy condition evaluation; Sonnet handles primary reasoning tasks.
The human-in-the-loop model is enforced at the code level through the Permissions primitive. Before any order reaches the broker, the agent generates a structured approval card showing the exact trade, every signal that triggered the recommendation, and the full reasoning chain. Users can choose between three consent modes: in-chat for synchronous approval, async for queued approvals when they are away, and autonomous execution within guardrails defined upfront in the Playbook. The Learnings journal captures the agent's reasoning at the moment of each decision — not reconstructed afterward — providing a permanent audit trail the agent can reference in future sessions.
All data, including <a href="/news/2026-03-14-onecli-open-source-credential-vault-and-gateway-for-ai-agents-built-in-rust">API keys and Dhan credentials</a>, trade history, and Playbooks, is stored locally in ~/.vibetrade/ with no cloud dependency. The project is at an early v0.1 stage and is actively seeking user feedback via its founders' Calendly.