EPO Publication Evidence MCP Server

Retrieve official European patent publications — by publication day, or by publication number — from the EPO publication server, live, inside the source's stated fair-use limits.

Built for patent research.

Fair use is enforced in code, not promised in a README

The EPO publishes a 10 GB per 7 days per IP fair-use allowance and links it from its own robots.txt. This server counts every byte over a rolling 7-day window and refuses a read that would cross a self-imposed ceiling of 1 GB — a tenth of the charter — kept deliberately small because this IP is shared with our EPO watch products.

A refusal shows its arithmetic, fetches nothing, and charges nothing. Every answer reports how much of the budget is spent and how much remains.

The arithmetic that makes this comfortable:

one publication day   ~506 KB in ONE request, covering every publication that week
one front page        ~13 KB

The kind code lives inside the publication id, so classifying a whole day of 6,241 publications costs no extra reads at all. This server never fetches a document it was not asked for.

Three things this server gets right that a naive read gets wrong

Measured live on 2026-07-30.

1. A bare patent number does not resolve — it is an HTTP 500

/patents/EP4780196NWA1/document.html  ->  200, 13,292 B
/patents/EP4780196/document.html      ->  500
/patents/EP4780196A1/document.html    ->  500
/patents/4780196/document.html        ->  500

The middle token (NW, W1) is not derivable from the number — it is only discoverable from a publication-day index. So if you hold a bare number, give it together with the publication day and this server resolves the id for you. If more than one publication carries that number on that day, you get the candidates, never a guess.

2. The middle token is not always letters — and the obvious fix is worse

Corrections arrive as EP4395092W1A8 — a digit in the middle token. On 2026-07-29, 39 of 6,241 ids carried one, and they are exactly the A8/B8/B9 correction slice. A parser written for [A-Z]{2} drops all 39 while appearing to work.

But loosening it to [A-Z0-9]{2} is wrong in the more dangerous direction — it makes EP4780196A1 match as number 47801, token 96, kind A1, accepting an id the source rejects and decomposing it into the wrong patent number. Measured over 12,008 ids across two weeks, the token alphabet is NW (11,922), W1 (85), W2 (1) — always letter-first. The pattern used here is [A-Z][A-Z0-9], and any id that still fails to parse is counted and reported, never silently skipped.

3. There is no cheap metadata endpoint

/bibliographic-data and /document.pdf both return HTTP 500. Only document.html serves — and it is parsed here into INID-coded fields (application number, filing date, IPC and CPC classes, PCT numbers, designated states) rather than handed back as a blob.

Tools

tool	what it answers
list_publication_days	The EPO's weekly publication days, newest first — the archive runs back to 1978-12-20 (2,424 weeks)
get_publications_for_day	Every EP publication issued on one day, filterable by kind code or number prefix, with kind-code counts
get_publication	One publication's INID-coded front page — by full id, or by number plus the day it published
explain_limits	The fair-use budget, the three traps above, and the kind-code table. Read before acting.

Scope

EP publications only, from the EPO publication server. There is no full-text search on this source — this is retrieval by day or by number, not a patent search engine. National patents, legal status and register events are outside it.

robots.txt on the publication server is User-agent: * with an empty Disallow — everything is permitted — followed by a pointer to the EPO terms of use and fair-use charter. That charter is why the budget above exists. Every call is additionally paced through one strict-pacing bucket at 2 requests/second.

This server retrieves what the EPO published. It does not assess novelty, validity, freedom to operate or infringement, and it is not legal advice.

Pricing

Event	Price
Server start (apify-actor-start)	$0.05, charged once per server run by the platform
Tool call (mcp-tool-call)	$0.05 flat per completed tool call

No other events. Memory is pinned to 1024 MB so the start fee is charged once — on this platform the start event bills once per gigabyte.

Errors and notices are not charged. A malformed publication id, a malformed publication day, an unknown kind code, an EPO read failure — and a fair-use refusal — all return without billing. If this server declines to spend more of a shared allowance, it fetched nothing and it bills nothing.

A sourced "nothing published under that number on that day" IS charged, and so is a day whose filter matches nothing — those are the answers to the questions asked.

If a tool call cannot be billed correctly, the call fails rather than serving free.