LIVE TODAY
An expert for one field. It answers from vetted sources with a citation, or refuses for free. Connects to your agent over MCP.
Examiner procedure from the USPTO's own manual, 8 core chapters. Every answer shows the exact MPEP passage it came from.
$0.10 / cited answer PREMIUM
$0.00 / refusal
Applied AI engineering from AI Engineer conference talks: agents, evals, retrieval, production LLM systems. Every answer shows its source.
$0.02 / cited answer
$0.00 / refusal
Startup guidance from YC partner talks and founder lectures: fundraising, product-market fit, growth. Every answer shows its source.
$0.02 / cited answer
$0.00 / refusal
Planned source: IRS publications (public record)
Get notified →Planned source: USCIS policy manual (public record)
Get notified →Planned source: FAA regulations (public record)
Get notified →Planned source: SEC rules & guidance (public record)
Get notified →Planned source: OSHA standards (public record)
Get notified →Planned source: EPA regulations (public record)
Get notified →WHAT IS AN EXPERT PACK?
An expert pack gives your AI agent the knowledge of someone who has spent years in one field. Your agent consults it before it answers, so you get a real domain expert's answer with the exact source attached. Not a confident guess.
Your agent asks a question it has to get right.
It answers only when the field's source material clearly backs it up.
It answers with the source attached, or it says it doesn't know. No source, no answer, and a "don't know" costs nothing.
Retrieval and evidence checks by DEX; answers generated by Claude over the pack corpus. Median 8.4s on the 18-question eval (2026-07); the full per-question log ships with the beta docs.
A trap question is one the sources cannot answer; the eval plants three, and a passing pack refuses all three.
FROM THE FOUNDER
DEX is built by one founder. No user reviews yet. The ★ row means a pack passed our published eval, not user opinion. Instead of reviews: nine recorded sessions and an 18-question eval where all 3 trap questions refused at $0. Questions? hello@dexforagents.com
HOW IT WORKS
Here is what happens when it does.
ENDPOINT FORMAT · YOURS IS ISSUED WITH YOUR BETA KEY
https://mcp.dexforagents.com/{your-pack}
Your agent queries the pack over MCP, mid-task, like any other tool call.
Every answer returns with a citation and chunk key to the primary source. If the sources can't back an answer, the pack refuses and refusals cost nothing.
MCP is the one-line way agents plug into tools. Claude Code, and most agent frameworks, already speak it.
Endpoints are issued per beta batch. There is no public URL yet.
Get a beta keyWHERE YOUR QUERIES GO
Your agent's queries go to the pack server, which retrieves from the corpus and generates the answer through Anthropic's API. We log queries, retrieved chunks, and answers to debug retrieval and run evals. We don't use your data to train models and we don't sell or share it.
TESTED CLIENT
Works with Claude Code. MCP is an open protocol; other clients are untested until we say otherwise. Beta batches are onboarded by hand, so bring the client you actually use.
Come with a question your agent keeps getting wrong.
Join the beta