What it does
Dexter is an open-source autonomous agent for financial research. It turns a complex financial question into a step-by-step research plan, chooses and runs tools to collect data, reviews its own results, and keeps refining until it has an answer backed by the data. It works with real-time financial data, including income, balance sheet and cash flow statements, and guards against runaway execution with loop detection and a cap on steps. Setup is local. You need the Bun runtime, an OpenAI API key and a Financial Datasets API key, with an optional Exa key for web search. Other model providers (Anthropic, Google, xAI, OpenRouter) and a local Ollama server can be configured in the environment file. You clone the repository, install with Bun and start it in interactive mode in a terminal. The repository ships with an eval suite that has an LLM judge score the agent on a set of financial questions, with results tracked in LangSmith. Each query's tool calls and reasoning steps are written to a JSONL scratchpad for debugging. A WhatsApp gateway lets you message the agent from your own chat. The project is MIT licensed. The authors say it is meant for learning and information only, not for actual trading or investing, and that its results can be wrong, partial or stale.
Capabilities
- Breaks financial questions into structured research steps
- Self-review loop with runaway-loop guards and a cap on steps
- Real-time income, balance sheet and cash flow statement data
- Configurable for OpenAI, Anthropic, Google, xAI, OpenRouter or local Ollama models
- Built-in eval suite and JSONL scratchpad logs of every tool call
- WhatsApp gateway for messaging the agent
When to use it
- You want to experiment with an agent that researches company financials from the terminal
- You want an open, modifiable starting point for a finance research agent
- You are studying how planning and self-checking agents behave on a data-heavy task
