virattt/dexter
An autonomous agent for deep financial research
Autonomous trading bots run on vibes. This one runs on filings.
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
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.
Crypto Twitter has a bot problem, and it's not the spam kind. Every other week there's a new "autonomous trading agent" thread — some wallet hooked up to an LLM, making buy/sell calls off sentiment scores and price momentum. Most of these get quietly deleted after a drawdown.
Here's the thing nobody wants to say out loud: an LLM reacting to price action in real time isn't doing anything a stop-loss order couldn't do cheaper. Where AI actually earns its keep in this space is the boring part — reading. Filings, earnings calls, 10-Ks, transcripts. The stuff that takes a human analyst three hours and that most retail traders skip entirely.
That's the job dexter does. It's an open-source financial research agent — 24.6k GitHub stars — that ingests company filings and earnings call transcripts and turns them into a structured research memo. No trade execution, no wallet permissions, no "vibes." It reads, it summarizes, it flags what matters, and it hands you a document before you size a position.
In liquid large-caps, reflexes and speed can actually win — you're competing against other fast actors in a market with real price discovery. Illiquid alts are a different game. Thin order books mean your entry and exit price is set by whoever else is paying attention, and most of CT isn't reading anything past the chart.
A trading bot optimized for reflexes is solving the wrong problem in an illiquid market. There's no edge in being 200ms faster than the next guy when the real edge is knowing something the next guy didn't bother to read. Research beats reflexes when liquidity is the constraint, not latency.
This is also where a lot of "DD" in crypto stops short. People screenshot a tokenomics chart, skim a Medium post, call it due diligence. For actual equities-adjacent plays — treasury companies, tokenized RWA issuers, anything with SEC filings or an earnings call attached — there's a whole layer of primary-source information that basically nobody in crypto reads because it's tedious. That's exactly the tedium an agent should be absorbing.
Say you're looking at a mid-cap company that's added a crypto treasury strategy — the kind of ticker that gets passed around in DeFi Discord servers as a "proxy play." Before you touch size, you want to know what management actually said, not what the thread said they said.
Set up dexter locally. Clone the repo, install dependencies (pip install -r requirements.txt in the standard setup), and add your LLM provider API key to a .env file — dexter works with OpenAI or Anthropic keys depending on your config.
Point it at the target. Feed it the company's ticker or the specific filing/transcript you want analyzed — most recent 10-Q and the latest earnings call transcript are the minimum useful set.
Ask a scoped question, not a vague one. Don't prompt "tell me about this company." Prompt something like: "Summarize management's stated crypto treasury strategy from the Q2 earnings call, flag any changes in guidance from the prior quarter, and list every risk factor mentioning digital assets in the latest 10-Q."
Let it generate the memo. Dexter pulls the source documents, extracts the relevant sections, and returns a structured summary — direct quotes tied to specific filings, not paraphrased vibes.
Cross-check the flagged risk factors against your thesis. If the memo surfaces a risk factor about "potential dilution from at-the-market equity offerings to fund treasury purchases" and your whole thesis is "they're never diluting," that's your signal to resize before you enter, not after.
Total time: maybe 20 minutes, most of it hands-off while the agent processes. Compare that to the alternative — skimming a 40-page 10-Q at midnight because someone tweeted "bullish" — and it's not close.
| | Research agent (dexter) | Autonomous trading bot | |---|---|---| | What it optimizes for | Information quality before you decide | Speed of execution after a signal | | Best market condition | Illiquid, thinly-traded, filing-heavy assets | Liquid markets with fast, clean price action | | Failure mode | Slow — won't catch a 30-second pump | Executes confidently on noise, no fundamental grounding | | Wallet/custody risk | None — it doesn't touch funds | Full custody risk if it has trade execution | | Where it adds edge | Reading what others won't | Nowhere it doesn't already have infra-level speed advantage |
If you're scalping liquid perps, a research agent won't help you — that's a latency game, not a reading game. But if you're evaluating a position you plan to hold for weeks or months, especially anything with real-world filings attached, a bot that trades on your behalf is the wrong tool entirely. It's solving "how fast can I act" when your actual problem is "do I even understand what I'm buying."
Treating the memo as a buy signal instead of an input. Dexter tells you what's in the filings. It doesn't tell you what to do about it. Feed the output into your own thesis, don't outsource the decision.
Feeding it stale or incomplete source documents. If you only give it last year's 10-K and skip the most recent earnings call, the memo will confidently miss the update that actually matters. Garbage in, confident garbage out — same as any LLM.
Using vague prompts and accepting vague output. "Summarize this company" gets you a Wikipedia-tier summary. Scoped prompts — specific quarters, specific risk categories, specific comparisons to prior guidance — get you something you can actually act on.
Skipping it because the target "doesn't have real filings." Fair — dexter's edge is filings and transcripts, so it's less useful for a pure on-chain-only token with no TradFi paper trail. In that case you're back to on-chain data tools, not a research agent. Know which problem you're actually solving before picking the tool.
No. It's a research and analysis agent — it reads filings and earnings calls and produces summaries. It has no wallet integration or trade execution capability, which is by design.
Less so. Its core strength is parsing structured financial documents — 10-Ks, 10-Qs, earnings call transcripts. For a token with no TradFi paper trail, you'll get more value from on-chain analytics tools than a filings-focused research agent.
The difference is structure and sourcing. Dexter is built specifically to ingest and parse financial documents at scale, tie summaries back to specific source passages, and handle multi-document context (comparing this quarter's call to last quarter's, for example) — rather than a one-off copy-paste summary.
→ 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.