virattt/dexter
An autonomous agent for deep financial research
Most research agents are GPT wrappers. One actually digs.
Turn what you learned into a concrete stack decision.
Want the shortlist in your inbox?
Subscribe for the weekly brief that turns new AI noise into the few tools and workflows worth testing.
An autonomous agent for deep financial research
Guide
Dexter: The Autonomous Research Agent Crypto Diligence Actually Needs
An open-source agent that pulls financials, cross-checks sources, and ships a report while you do something else.
Guide
The AI Stack for Web3 Builders: Research, Agents, and Onchain Payments
The full map: which AI tools handle DD, trading analysis, orchestration, and payments.
Guide
Dexter: The AI Research Agent DeFi Traders Aren't Using Yet
23k stars, zero hype. This autonomous research agent does the DD most traders skip.
Two weeks ago I set out to answer a boring but important question: if you're building in crypto and need to research a token, a protocol, or a macro thesis, which AI agent actually does the work instead of just summarizing whatever's in its training data?
I ran the same set of prompts against a dozen "AI research agent" repos and products. Most of them are a system prompt, a search API call, and a landing page. One survived: Dexter, an open-source autonomous financial research agent with 24.7k GitHub stars. Here's what separated it from the pile, and how to actually use it.
Type "AI research agent" into GitHub and you'll get hundreds of results. Almost all of them follow the same pattern:
That's a one-shot summarizer wearing a research agent costume. It works fine for "what is restaking" — questions Wikipedia already answers. It falls apart the moment you ask something that requires pulling from multiple sources, cross-checking numbers, and reasoning about whether they agree.
I tested this directly: I asked each tool the same question — "compare protocol revenue trends for a specific L2 across the last two quarters, and flag if TVL growth is organic or incentive-driven." Nine out of twelve tools gave me an answer that read well but was built on a single blog post or a stale API snapshot. Two hallucinated numbers outright. Only Dexter actually pulled multiple sources, cross-referenced figures, and told me when data was thin instead of papering over it.
Dexter isn't a chat wrapper — it's built around a multi-step research loop: plan the investigation, pull from multiple financial data sources, cross-check figures against each other, and only then synthesize an answer. That loop is the whole point. It's the difference between "here's what one article said" and "here's what three sources say, and here's where they disagree."
A few things stood out over two weeks of daily use:
It's not magic — it's slower than a one-shot summarizer because it's actually doing multiple retrieval passes. If you want a fast, shallow answer, this is the wrong tool. If you want something you'd actually put in front of an investment committee or a DAO proposal, it's the only one of the twelve that didn't embarrass itself.
Here's the exact test I ran, step by step, so you can replicate it.
The setup: I picked a mid-cap L2 token and wanted to answer: "Is this token's recent price action supported by fundamentals, or is it narrative-driven?"
Step 1 — Give it the thesis, not just a ticker. I didn't ask "tell me about [token]." I gave it a specific claim to test:
Thesis: [Token]'s 40% price increase over the last 30 days is driven by
real usage growth (fee revenue, active addresses), not just a narrative
pump. Verify this against on-chain and financial data. Flag any metric
that doesn't support the thesis.
Step 2 — Watch the research plan. Dexter broke this into sub-questions before pulling anything: fee revenue trend, active address trend, token unlock schedule, and whether volume was concentrated on a few exchanges (a red flag for wash trading or low organic interest).
Step 3 — Review the source-by-source pull. It pulled fee revenue data from one source, active address counts from a block explorer-style source, and cross-referenced both against the price chart timeframe. Fee revenue was up 12% over 30 days — real, but nowhere near the 40% price move.
Step 4 — Get the verdict, with the gap called out explicitly. The output didn't say "yes, fundamentals support it" or "no, it's a pump." It said: usage growth is real but modest; the price move outpaces usage growth by roughly 3x; check the unlock schedule and exchange volume concentration before treating this as a fundamentals-driven move. That's a genuinely useful, falsifiable answer — not a vibes-based one.
Step 5 — Follow up on the weak point. I asked it to dig specifically into exchange volume concentration. It re-ran just that leg and came back with the breakdown instead of re-doing the whole analysis. That's the iteration loop working as intended.
Total time: about 6 minutes for the full pass, versus maybe 90 seconds for a one-shot summarizer — and the one-shot tools I tested with the same prompt just said "yes, growing usage supports the price action" with no numbers attached.
| Situation | Use Dexter | Skip it | |---|---|---| | Due diligence on a token before a treasury allocation | Yes | — | | Writing a DAO proposal that needs defensible numbers | Yes | — | | "What's the current price of X" | — | Use any price API, this is overkill | | Quick sentiment check on Crypto Twitter | — | Use a social listening tool instead | | Comparing protocol revenue across a quarter | Yes | — | | You need an answer in under 30 seconds | — | One-shot summarizer is fine here |
The honest tradeoff: Dexter is overkill for casual questions and underkill for nothing — it's built for the exact use case of "I need to trust this number enough to act on it." If your research bar is "sounds plausible," a GPT wrapper with a nice UI will feel faster and you won't notice the gap until you're wrong.
Treating it like a chatbot instead of giving it a falsifiable thesis. "Tell me about this protocol" gets you a generic summary even from a good research agent. "Verify this specific claim against data" is what triggers the actual multi-source research loop. Feed it a hypothesis, not a topic.
Not reading the source breakdown. The temptation is to skip straight to the summary paragraph. The actual value is in seeing which sources contributed what — that's how you catch a conclusion that's secretly resting on one thin data point.
Running it once and treating the answer as final. The iteration loop is the point. If a number looks off or a source seems weak, push back and let it re-run that specific piece. People run it once, get a plausible-sounding answer, and stop — that defeats the reason to use a multi-step agent over a one-shot summarizer in the first place.
Using it for things that don't need it. If you just need a current price or a quick TVL number, you're paying a latency and compute tax for a research loop you don't need. Save it for questions where being wrong actually costs you something.
An autonomous research agent plans multiple sub-questions, pulls from more than one source per question, and cross-checks the results before synthesizing an answer — often with the ability to re-run a specific step based on follow-up. A chatbot with search does one query and summarizes the top hits in a single pass.
It's open source, so the code itself is free — you're running it yourself and covering whatever API/data source costs it hits. Check the tool page for current setup details before committing time to it.
Treat any AI agent's output as a starting point for verification, not a final answer — even a good one like Dexter will tell you when data is thin rather than hide it, which is exactly the signal you should listen to. Use it to surface the numbers and disagreements fast, then verify the specific claims that matter before you act on them.
→ Ask the index what to build your financial research stack
→ Free credits for these tools
Written by McKlaud AI. Want to know which AI tools actually fit your business? Get a free AI audit.