Dexter

About

Dexter is an autonomous AI agent that performs deep financial research with a level of rigor most paid tools can't match — and it's completely free and open source. Give it a question like "Compare NVIDIA's cash flow trajectory against AMD over the last 8 quarters" and it breaks the problem into discrete research tasks, pulls live income statements, balance sheets, and SEC filings, then validates its own conclusions before presenting them to you. The architecture is what makes Dexter different from throwing a financial question at ChatGPT. Four specialized agents work in sequence: a Planning Agent decomposes your query into steps, an Action Agent executes each step using real financial data tools, a Validation Agent checks the results for consistency and accuracy, and an Answer Agent synthesizes everything into a coherent analysis. If validation fails, the system loops back and re-researches — you never see unverified numbers. Dexter runs entirely in your terminal via Bun runtime. It connects to Financial Datasets API for real-time market data, supports six LLM providers (OpenAI, Anthropic, Google, xAI, OpenRouter, and Ollama for fully local execution), and includes an evaluation suite that benchmarks the agent against known financial questions using LangSmith. The scratchpad feature logs every tool call to JSONL files in .dexter/scratchpad/, giving you full transparency into how the agent reached its conclusions. You can trace exactly which data sources it queried, what it found, and why it chose to re-validate certain claims. For solo investors, the pitch is simple: you get institutional-grade research workflows without paying Bloomberg Terminal prices. For developers, Dexter is a reference implementation of a self-validating multi-agent architecture in TypeScript that's clean enough to learn from. The MIT license means you can fork it, extend it, and build commercial products on top. The honest limitation: Dexter is a CLI tool. There's no web dashboard, no pretty charts, no portfolio tracking. You need API keys for Financial Datasets (and your preferred LLM provider), and the quality of analysis depends heavily on which LLM you use. OpenAI is the primary supported provider, and switching to cheaper models noticeably degrades output quality on complex multi-step queries.

Key Features

• Four-agent architecture: Planning, Action, Validation, and Answer agents work in sequence
• Self-validation loop catches errors before presenting results — re-researches when findings are inconsistent
• Live financial data: income statements, balance sheets, cash flow, and SEC filings via Financial Datasets API
• Six LLM providers supported: OpenAI, Anthropic, Google, xAI, OpenRouter, and Ollama (fully local)
• Scratchpad debugging: every tool call logged to JSONL files for full transparency
• Built-in evaluation suite with LangSmith integration and LLM-as-judge scoring
• Loop detection and step limits prevent runaway execution and cost overruns
• WhatsApp gateway for receiving financial research results on your phone

Use Cases

• 1Compare quarterly financials across competitors — ask 'How does NVIDIA's gross margin trend compare to AMD's over the last 2 years' and get a validated analysis
• 2Research SEC filings without reading 200-page documents — Dexter extracts the specific data points you need
• 3Build a research workflow that runs locally with Ollama — no API costs, no data leaving your machine
• 4Evaluate LLM accuracy on financial questions using the built-in eval suite before trusting production outputs