Open-source Claude Code agent for stock market, 26k stars on github

追加: 2026-05-27

[00:00:00]Someone built a finance agent that's basically Claude code in a finance hat.

[00:00:04]26,000 GitHub stars in 7 months. The whole thing is one TypeScript file. So, what's actually inside it? Let's take a look.

[00:00:13]You type a question.

[00:00:15]What was Apple's gross margin trend last four quarters? It calls a tool, hits a financial data API, pulls earnings, writes you an answer.

[00:00:25]The readme says it out loud. It's Claude code pointed at finance. Same shape, autonomous agent, terminal first, tool using, different domain. And the interesting part isn't the answers, it's that the entire agent is small enough to read in an afternoon. For 2 years, every LLM lived inside a chat box. Then Claude code put one in a terminal. Then cursor put one in your editor. Now they're leaking into actual domains, legal research, genomics, ops. Dexter is the worked example for Claude code, but for finance. And the value isn't the finance part, it's that you can read the template. Here's the whole architecture, one agent file, about 680 lines, a single class, and a plain while loop.

[00:01:12]User question goes in.

[00:01:15]The LLM responds. If the response contains tool calls, run them. Push results back. Loop. If it doesn't, that's the answer. Return. No LangGraph, no DAG, no state machine. The agent framework industry spent 2 years abstracting this. Dexter just wrote the loop.

[00:01:33]The readme brags about intelligent task planning, automatic query decomposition into structured research steps. So, grep the repo for a planner. Zero hits.

[00:01:44]There's no planner file.

[00:01:46]The planning is just the system prompt plus LangChain's bind tools.

[00:01:50]Decomposition is whatever the model decides to do, and the prompt actively discourages chopping things up. Do not break up queries.

[00:01:59]The intelligent planning is the LLM being told not to plan. Tools live in one file, 238 lines, a declarative array. Each entry carries a name, a description, and a concurrency safe flag. That last bit matters. The flag lets the agent run safe tools in parallel. Financial data lookups, web searches, all firing at once. 14 tools, one array, no decorator magic. The cleanest tool registry I've read in a while. Next big claim.

[00:02:29]Self-validation.

[00:02:31]It checks its own work and iterates until tasks are complete. Open the scratchpad file. The self-validation is 30 lines of Jaccard similarity. Before each tool call, it compares the new query against the last three calls to the same tool. If the word overlap is too high, it injects a warning into the next prompt. Three attempts per tool.

[00:02:53]Then it nags harder. Critically, it never blocks a call. It just complains.

[00:03:00]It's not an LLM critic. It's a deterministic loop detector with delusions of grandeur.

[00:03:06]And honestly, it works. Every agent tutorial ducks this question. How does the loop end? Dexter's answer is one line. When the LLM stops asking for tools, the last text is the answer.

[00:03:20]That's it.

[00:03:21]Hard cap, 10 iterations.

[00:03:23]After that, you get reached maximum iterations. I was unable to complete the research.

[00:03:31]One if statement and a counter, done.

[00:03:33]Eight LLM providers out of the box, every name you'd want.

[00:03:38]But, open the providers file. 100 lines, one array, routing by model name prefix, and four of those providers just reuse the OpenAI SDK with a different base URL.

[00:03:50]So, eight providers is really four SDKs and a dictionary. Boring, effective. The part nobody mentions, Dexter ships a WhatsApp gateway.

[00:04:00]16 files built on Baileys, the unofficial reverse-engineered WhatsApp library.

[00:04:07]Scan a QR code, pair your phone, and the agent replies to messages. The default UX is message yourself. Genius and slightly cursed. Here's the punchline, only three things in this repo are finance-specific.

[00:04:21]The finance tools. The identity prompt that tells the agent it's a financial researcher. And the compaction prompt that speaks analyst. Swap those three.

[00:04:32]Keep the loop, the registry, the providers, the WhatsApp gateway.

[00:04:36]You now have Claude code for legal research, for genomics, for your domain.

[00:04:42]Reality check.

[00:04:43]The readme claims an MIT license.

[00:04:45]There's no license file committed. The default model is hard-coded to a name that doesn't ship in any catalog I can find. The compaction prompt literally threatens the LLM.

[00:04:56]Tool calls will be rejected. You will fail the task. About seven test files for tens of thousands of lines of code.

[00:05:04]This is a strong reference implementation, not a hardened product.

[00:05:08]Repo's linked below. Don't run it with real money. Steal the loop, throw away the finance. That's the template.