Dexter is an autonomous agent for financial research. Given a question, it decomposes the query into a structured research plan, selects and runs tools to gather live market data, checks its own results, and iterates until the task is complete. The financial tools return income statements, balance sheets, and cash flow statements; a web-search tool is optional. Loop detection and step limits bound execution.
It runs on the Bun runtime as an interactive terminal session. Required credentials are an OpenAI key and a Financial Datasets key; Exa (with Tavily as a fallback) enables web search, and Anthropic, Google, xAI, OpenRouter, or a local Ollama endpoint can be configured as alternative model providers. A WhatsApp gateway links your phone so that messages you send to yourself are answered by the agent in the same chat.
Every tool call is logged to a per-query JSONL scratchpad under .dexter/scratchpad/, recording the original query, each tool invocation with its arguments, raw result and LLM summary, and the agent's reasoning steps. An evaluation suite runs the agent against a dataset of financial questions, scores correctness with an LLM-as-judge approach, and reports to LangSmith, with a live view of progress and accuracy. The project describes itself as intended for educational and informational use rather than trading or investment.
Features
- Task planning: decomposes a complex question into ordered research steps
- Autonomous tool use: selects and executes tools to fetch the needed financial data
- Self-validation: checks its own work and iterates until the answer is complete
- Financial statements: income statements, balance sheets, and cash flow statements from the Financial Datasets API
- Web search: optional Exa search with Tavily as a fallback
- Provider choice: OpenAI by default, with Anthropic, Google, xAI, OpenRouter, or Ollama configurable
- Safety limits: loop detection and step limits to stop runaway execution
- Scratchpad logs: per-query JSONL records of every tool call, result, and reasoning step
- Evaluation suite: LLM-as-judge scoring over a question dataset, tracked in LangSmith
- WhatsApp gateway: chat with the agent from your phone by messaging yourself
