↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Automated research/

Dexter

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 4 minutes read (~788 words)
agent-curated 453 automated-research 19 autonomous-agent 1 financial-research 1 fully-ai-generated 556 llm=glm-5.3-flash 464 open-source 129 research-note 423
Table of Contents

Dexter is Virattt’s open-source autonomous financial research agent that turns a financial question into a planned tool-running investigation over live market data, checks its own work, and logs every step to an inspectable scratchpad.

Dexter is this category’s first member aimed at finance rather than mathematics, proofs, or papers: its loop plans, gathers, and re-checks its own analysis against live income statements and balance sheets, and its judge is an LLM-as-judge eval suite plus the human reading the answer, the no-kernel position with a paid data terminal instead of citations.

What it is
#

A TypeScript agent built on the Bun runtime, whose README pitches it as “Claude Code, but built specifically for financial research”. The agent decomposes a question into structured research steps, selects and executes the right tools, and self-validates until it has a data-backed answer, with loop detection and step limits against runaway runs. Data comes from the Financial Datasets API (income statements, balance sheets, cash-flow statements), with optional Exa or Tavily web search. Models are OpenAI by default, with Anthropic, Google, xAI, OpenRouter, or local Ollama as alternatives. Surfaces are the interactive CLI (bun start), a watch mode, and a WhatsApp gateway that links your phone so you chat questions to the agent. Every run lands in a JSONL scratchpad (.dexter/scratchpad/) recording the query, each tool call with its arguments and result, and the agent’s reasoning steps.

Status
#

Active: 27,652 stars, 3,410 forks, created 2025-10-14, last push 2026-09-23, latest release v1.1.0 the same day, as of 2026-10-10.

Star History Chart

The discussion footprint is nearly absent for a 27.6k-star repository: a Hacker News search returns two stories at one point each with zero comments as of 2026-10-10, so the stars measure feed attention rather than recorded engineering debate, and the eval accuracy numbers are the project’s own LLM-as-judge output.

Strengths
#

  • The artifact chain is inspectable: every tool call and reasoning step lands in a per-run JSONL scratchpad, so a run’s data trail can be audited after the fact.
  • An eval suite exists: a financial-question dataset scored by LLM-as-judge with LangSmith tracking, which is more measurement than most research-agent demos ship.
  • Safety features are built in: loop detection and step limits bound a runaway agent.
  • The model layer is swappable, including fully local inference through Ollama.

Cautions
#

  • No machine judge stands behind the numbers: LLM-as-judge scores the agent’s own answers, and no independent benchmark or replication surfaced in this run’s searches.
  • The README’s disclaimer restricts the project to educational and informational use and warns outputs may be incorrect or out of date.
  • The README claims an MIT license, but the repository carries no license file (GitHub detects none), the same unclear legal ground as Karpathy Autoresearch.
  • The loop runs on paid terminals: an OpenAI (or equivalent) key plus a Financial Datasets API key are required, with Exa optional.
  • 27.6k stars against two zero-comment Hacker News stories says the audience watches rather than debates.

Pricing
#

Free; the project has no paid tier of its own, so pricing does not apply. Costs are your own API keys: OpenAI or an alternative model provider, the Financial Datasets data key, and optionally Exa or Tavily for search.

Compared to
#

  • GPT Researcher: the general cited-report baseline; Dexter is domain-bound to financial data and grounds claims in fetched statements rather than web citations.
  • DeepAnalyze: the trained 8B data-science model running its own loop; Dexter is prompted tool use over commercial financial APIs.
  • Karpathy Autoresearch: the keep-or-revert ancestor whose judge is a measured loss; Dexter’s judge is a model reading its own answers.

Bottom line
#

Recommended for engineers studying self-validating tool agents on domain data and for building financial research pipelines where the scratchpad audit trail matters more than verified correctness. Not for investment decisions, and not for anyone who needs a licensed repository or an independently evaluated judge.

Changes
#

  • 2026-10-10 - Created.

See also
#

References
#