virattt/dexter
An autonomous agent for deep financial research
An open-source agent that pulls financials, cross-checks sources, and ships a report while you do something else.
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
I Tested AI Research Agents for Crypto Builders for 2 Weeks. Here's What Survived
Most research agents are GPT wrappers. One actually digs.
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.
Most people doing "research" on a crypto project are doing the same thing: 6 browser tabs, a half-filled spreadsheet, and a gut call at 11pm because they're tired of tab-switching. That's not diligence, that's vibes with extra steps.
Dexter is a different approach. It's an open-source autonomous research agent — 24.6k GitHub stars, built by the same person behind the popular ai-hedge-fund project — designed to do the actual digging: pull financial data, cross-reference multiple sources, catch contradictions, and hand you a structured report instead of a pile of open tabs. It was built for equities and financial research generally, but the workflow maps cleanly onto crypto diligence — treasury health, tokenomics claims, revenue numbers, whatever a project is telling you versus what the data says.
I ran it against a real project's numbers to see if it holds up. Short version: it's genuinely useful for the grunt work, not a replacement for judgment on the parts that actually matter.
Dexter isn't a chatbot you interrogate. It's an agent loop: you give it a research question, it plans a set of sub-tasks, executes them (pulling data, running searches, calling tools), evaluates whether the answer is actually supported, and either ships a report or goes back to dig more.
The core loop looks like this:
The part that matters for crypto specifically is step 3. Most AI research tools will happily summarize a project's own marketing copy back to you with a confident tone. Dexter is built to cross-check — pull the number from one source, pull it again from another, and flag when they disagree. That's the whole point of diligence and it's the part people skip when they're doing it manually at midnight.
It's self-hosted and code-first — you're running it, not subscribing to a hosted dashboard. That means setup friction, but also means you're not trusting a black box with your research process.
Here's an actual walkthrough, not hand-waving. Say you want to sanity-check a mid-cap DeFi protocol's claim that it holds "$40M in protocol-owned treasury, fully diversified."
Setup:
git clone https://github.com/virattt/dexter
cd dexter
uv sync
cp .env.example .env
You'll need to drop in your API keys in .env — an LLM key (OpenAI or Anthropic) and whatever data source keys the tools you're using require. For crypto work, that typically means wiring in a blockchain data source (Dune, DeFiLlama's API, or a block explorer API) since Dexter ships configured for traditional financial data by default — you're extending its toolset, not using it out of the box for on-chain data.
Run it:
uv run dexter
The prompt you'd give it:
"Research [Protocol]'s treasury composition. The team claims $40M in protocol-owned treasury, fully diversified across at least 5 assets. Verify this against on-chain treasury wallet data and the protocol's own reporting (governance forum, docs). Flag any discrepancy between claimed and actual composition, and note if 'diversified' is doing any heavy lifting (e.g., is it 90% their own governance token relabeled as treasury?)."
What happens next: Dexter breaks this into sub-tasks — identify the treasury wallet address(es), pull current holdings, cross-reference against the team's public statements, check the diversification claim against actual asset composition. It executes each step, and critically, it doesn't just report "$40M confirmed" — a well-run pass will surface something like "$40M total, but $31M of it is the protocol's own native token at current market price, not external assets. 3 assets, not 5+." That's the discrepancy you were actually trying to find, and it's exactly the kind of detail that gets lost when you're skimming a governance forum post at speed.
Time cost: a manual version of this — finding the wallet, checking a block explorer, reading the forum post, doing the math on token concentration — is realistically 30-40 minutes done properly, and most people don't do it properly because it's tedious. Dexter's agent loop runs in the background while you do something else, and gives you a report you can then spend 5 minutes fact-checking instead of 40 minutes assembling.
The obvious alternative is pasting a project's whitepaper into a chatbot and asking "does this check out?" Here's when each makes sense.
| | Dexter | ChatGPT / Claude directly | |---|---|---| | Pulls live data itself | Yes — connects to real data sources | No — only knows what you paste in or what's in training data | | Cross-checks claims against sources | Built into the loop (verify step) | Only if you manually ask it to and feed it the sources | | Setup effort | Real — clone repo, API keys, tool config | Zero — open a browser tab | | Best for | Recurring diligence, treasury/financial verification, anything where "trust but verify" matters | Quick gut-check, explaining a concept, one-off questions | | Output | Structured report with citations | Freeform text, citations only if prompted | | Cost | Your own LLM API usage + data source costs | Subscription or free tier |
If you're doing diligence once on one project, ChatGPT with good prompting and manually pasted sources gets you 70% of the way there for zero setup cost. If you're doing this repeatedly — evaluating multiple protocols, tracking treasury health over time, running the same verification pattern weekly — Dexter's agent loop pays for the setup time fast because it's not re-explaining itself every session, and it's actually pulling fresh data instead of working off whatever you remembered to paste.
Skip Dexter entirely if you just want a vibe check on a project you're not seriously considering. It's overkill for "should I look into this more" — save it for "I'm about to put real money in, verify the numbers."
Trusting the report without checking the sources it cites. Dexter cites its sources, which is great — but an agent can still misread a document or grab a stale number. Treat the report as a massively accelerated first draft, not a signed-off audit. Click through on anything that changes your decision.
Not customizing the tool config for on-chain data. Dexter's default setup leans traditional-finance. If you don't wire in a blockchain data source, it'll happily research the wrong thing or fall back to general web search, which is much weaker than a direct on-chain query. Set up the integration before you run anything that matters.
Asking vague questions. "Research this project" gets you a shallow, generic report. "Verify the treasury diversification claim against on-chain wallet data and flag if the team's own token is being counted as diversified assets" gets you the actual answer. Specificity in, specificity out — same as any agent.
Running it once and calling it done. The value compounds when you use it as a recurring check — treasury composition, holder concentration, revenue claims — not a one-time report you file away. Numbers move; a report from three months ago tells you what was true three months ago.
Dexter itself is open-source and free. Your costs are the LLM API calls it makes (OpenAI/Anthropic usage) and any paid data source APIs you connect for on-chain data. For occasional research runs this is usually a few dollars, not a subscription-scale cost.
Not fully out of the box — it ships configured primarily for traditional financial research. You'll need to connect a blockchain data source (DeFiLlama, Dune, or a block explorer API) as a tool for it to pull wallet balances, treasury data, or token flows directly.
Dashboards show you data; Dexter does the reasoning on top of it — pulling from multiple sources, checking whether claims match reality, and writing up findings in plain language. It's not a replacement for those data sources, it's an agent that uses them as inputs to produce an actual verified answer to your specific question.
→ Ask the index what to build your financial research agent 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.