European law API for AI agents: Pandex over MCP | MewCP
MCP ecosystem
Pandex puts 2.8M European legal documents behind an MCP server
by Rohit Gite, Co-Founder @MewCP··5 min read
FAQ
Frequently asked questions
01
What is Pandex?
[+][−]
Pandex is a REST API and MCP server for official European legislation and highest-court decisions, covering the EU, seven European countries and the European Court of Human Rights.
While in beta, the Free plan is €0 for 1,000 requests a day. Heavier use costs €1 per 1,000 requests.
03
Does Pandex cover every EU country?
[+][−]
No. It covers EU-level law plus France, Germany, the Netherlands, Austria and Luxembourg, along with the UK and Switzerland. Eight more countries are listed as next.
Coding agents
OpenComputer is an open source Devin alternative you can self-host
OpenComputer runs Claude Code, Codex or OpenCode on a VM in your own cloud and takes tasks from Slack. We look at its speed benchmark, its single-VM design and what Hacker News asked.
Pandex serves legislation and top-court rulings from the EU and seven European countries through a REST API and an MCP server. Its coverage register, which lists its own gaps, is the most useful part.
No, the EU did not pass a law about MCP. someone put Europe's laws behind an MCP server instead, and ngl that's way more useful.
Pandex launched on Hacker News on Saturday with a simple pitch: AI models get European law wrong, "quoting old versions or making up text," so give them the official text instead. It got 103 points and 26 comments, and at least one commenter wondered why the EU had introduced an MCP law. Fair tbh.
My take: this is the right kind of boring. It won't turn your agent into a lawyer, but it fixes wrong citations and outdated versions, the exact failure that gets legal AI demos laughed out of the room.
ok so what did they actually ship
Per the Pandex homepage, it's legislation and highest-court decisions from the EU and seven European countries, pulled "straight from the institutions that publish them." The headline numbers are 2.8M official documents (2.4M court decisions and 421K acts and codes), 112M searchable passages, 37M citations linked to their text, and 33 of 41 institutions held complete.
The seven countries are France, Germany, the Netherlands, the UK, Austria, Switzerland and Luxembourg, plus the European Court of Human Rights. Eight more are listed as next.
You get it as a REST API or an MCP server, in Markdown or JSON, with keyword, semantic or hybrid search in any language. The homepage example asks in English whether a bike courier working for a delivery app is an employee and pulls up a 2018 Cour de cassation ruling in French. Every law comes with its past versions and the date each one took effect, and the maker says it's updated every night.
Setup is one line no key needed:
claude mcp add --transport http pandex https://api.pandex.io/mcp
Pricing is lowkey generous. Free is €0 while in beta with 1,000 requests a day, Usage is €1 per 1,000 requests, and a Feed plan from €1,000 a month hands you the whole corpus as files.
the best part is the page admitting what's missing
hot take: the coverage register is the real product here.
Most data APIs flash a big number and hope you don't ask. Pandex lists every institution, what it holds and what it doesn't, and why. France's Cour de cassation is partial because decisions published only through Judilibre since 2020 need a registered account, and the dumps run about two months behind. German constitutional court chamber decisions from 1998 to 2009 sit on a site that "refuses clients that are not a browser." About 900 House of Lords judgments aren't held at all.
And Luxembourg's Cour de cassation keeps its older judgments behind a search that asks for a proof-of-work challenge. Kinda iconic for a court.
This matters more for agents than for humans. A lawyer knows the database has holes. Your agent doesn't, it's gonna happily say "no relevant case law found" when the truth is "the case law is behind a Cloudflare wall." If you're building on this read that register and put the gaps in your system prompt or your answer template fr.
the HN comments were not here to clap
Fans mostly asked for more countries (Sweden, Italy, Poland, and one person wants US states). Someone else noticed the page used a picture of the US Supreme Court for a European law product, and the reply was a perfect "You're absolutely right! Let me fix that for you."
The skeptics had real points tho. One argued that AI working with law can't run on raw APIs like this, and that what you want is an interface that links every claim back to the exact paragraph. A long comment described law as a tripod: the written text, the case law around it, and how courts apply it, then said it's hard to compete with LexisNexis.
Another asked about an ordinance that says "valid from 1st of January 2027" but gets promulgated after that date, which Switzerland and Austria handle differently. And one team said they run a similar internal system where each question takes 2 to 5 minutes and costs about 1 to 2 dollars in tokens, because injecting raw law books into an LLM isn't reliable enough.
my read: the skeptics are right that this isn't a legal reasoning engine and wrong that it's pointless. Pandex doesn't claim to interpret anything. It hands the model the actual text, the right version and the linked citations, which is the input every one of those smarter systems still needs.
The "should be offered by the EU itself" comment is also... correct? It just isn't.
if your agent touches law, do this
The obvious use is grounding. Make the agent quote the passage and the version date it got back, and don't let it answer without one. That's a validation step on your side, not a Pandex feature.
The less obvious one is compliance monitoring, which Pandex also pitches. Say your agent checks the articles you care about every morning and when a new version shows up it posts the change in Slack and opens an issue in Linear. Pandex is the easy half, no key. The Slack and Linear half is where you juggle OAuth tokens and pray the agent never deletes anything.
That's the part I'd route through MewCP: the keys sit in its vault instead of your config, and you can let that agent post and create issues but not delete. Its call log shows which tool ran, when, the status and latency, not what came back, so keep your own record of the cited paragraph.
stuff we still don't know
We found no published accuracy benchmark, so how often search surfaces the wrong article from the right law is anyone's guess. The page shows the date each version took effect, but nothing we saw explains how disputed in-force dates are handled.
Also unclear: what "free while in beta" turns into later, and who's behind it. The homepage doesn't name the team, and the coverage we found mostly recaps the launch post.