What it does Dexter takes a natural-language financial question, decomposes it into a research plan, and executes it against live income statements, balance sheets, and cash-flow data. It validates its own work, iterates when results look shaky, and caps itself with loop detection and step limits so it doesn’t spiral.

The interesting bit The debugging paper trail is unusually thorough: every query spawns a JSONL scratchpad logging the original prompt, each tool call with raw results, and the LLM’s own summary of what it thinks it found. That transparency is rarer than it should be in agent projects. There’s also a WhatsApp gateway that lets you message yourself and get answers back in the same chat—an odd but practical UI choice.

Key highlights

• Built on Bun, with support for OpenAI, Anthropic, Google, xAI, OpenRouter, or local Ollama
• Evaluation suite uses LangSmith and an LLM-as-judge scorer with a real-time accuracy UI
• Financial data comes from Financial Datasets API; web search via Exa (Tavily fallback)
• Scratchpad files live in .dexter/scratchpad/ for full auditability
• MIT licensed

Caveats

• Requires three separate API keys minimum to do anything useful (OpenAI, Financial Datasets, and optionally Exa)
• The README’s disclaimer is extensive: outputs may be wrong, incomplete, or stale, and the project is explicitly not for real trading
• No benchmarks or accuracy numbers are published; the eval framework exists, but performance claims are absent

Verdict Worth a look if you’re building agentic research tools and want a reference implementation with good observability. Skip it if you need production-grade financial advice or are allergic to API key juggling.