patent-fetch

Find existing patents, and keep a deduplicated, private library of the ones you read. Give it some words, or a publication number; it asks a set of swappable patent services, merges what they answer, and says by name which of them answered and which did not. A patent retrieved once is stored with its description, its claims and a provenance record, and is never retrieved again.

It ships as a Python library and a command line (patent-fetch). It runs no language model: searching, merging and bookkeeping are deterministic.

$ patent-fetch search "cochlear implant electrode array" -n 5
  europepmc        ok             5 hits
  epo-ops          skipped        0 hits  needs EPO_OPS_KEY, EPO_OPS_SECRET
  lens             skipped        0 hits  needs LENS_API_TOKEN

5 distinct patents
    US20110264183    2011 PRE-CURVED COCHLEAR IMPLANT ELECTRODE ARRAY  (GIBSON PETER) [europepmc]
    US7272449        2007 Cochlear implant electrode array  (DADD FYSH) [europepmc]
    ...
$ patent-fetch fetch US7861317B2
✓ US7861317B2  [retrieved]  2011  Nose cover
    source=google-patents  full text  status=Expired - Fee Related
$ patent-fetch fetch "US 7,861,317"                    # second time: no network at all
✓ US7861317B2  [library]  2011  Nose cover
$ patent-fetch claims US7861317
I claim:
1. A flexible nose cover configured to attach to eyewear and shield the nose from frostbite ...

What it is for, and what it is not

Two jobs: a one-off search (“is there a patent on this?”), and keeping track of the patents in an area you work in.

• A collection is a named area: the queries that define it, the patents you decided belong to it, and every patent its searches have ever shown. A sweep runs the queries again and reports only what was not shown before, so an area can be re-checked without re-reading it.

  $ patent-fetch collection hearing --create --query "cochlear implant electrode array"
  $ patent-fetch sweep hearing                 # everything is new the first time
  $ patent-fetch collect hearing US7272449     # this one matters to us
  $ patent-fetch sweep hearing                 # next month: only what has appeared since
  

• One entry per invention. An invention is usually published several times: an application, its grant, the same filing at other offices. Publications that share a priority are folded into one hit, with the others listed under it (also published as ...), so one patent family does not fill the results. --all-publications lists each separately.
• It is not a legal opinion. A search shows what the enabled providers matched for these words on this day. No hits is not evidence that no patent exists, and nothing here says whether you are free to operate, whether a patent is valid, or whether it is in force: those turn on the claims as a patent attorney reads them. A legal status, an assignee or a priority date is passed on in the provider’s own words, and several providers say of their own data that it is not a legal conclusion.
• Loud, per-provider failure, of two kinds. A provider that is down, rate-limited or out of allowance is reported by name as unavailable, the other providers answer, and it is left alone for a while. A provider that fails because of a defect (it rejected our request, or its answer changed shape) is reported as error, named in the search’s broken list and logged with its traceback (patent-fetch problems). Neither is ever folded into “no results”.
• Private. A patent is public, but a copy came from a service with terms of its own, and a record of what you searched for says what you are working on. Files are 0600 in directories 0700; S3 objects never get an ACL and are checked with an anonymous request.

Providers

europepmc, google-patents, epo-ops and serpapi have been run against their services. lens has not: it is written from the service’s documentation and marked NOT YET RUN LIVE by patent-fetch providers until someone with an account runs it.

With no account at all you can read any patent by its number, but you cannot search for recent ones: the only keyless search is an archive of life-science patents that ends in 2012. To search what exists today, in any field, you need at least one keyed provider. epo-ops is the one to get first: free, official and worldwide. Until then a sweep can never report anything new.

Google Patents has no API. Its robots.txt allows /patent/ pages and disallows everything else, including its search. So google-patents requests one document page per lookup, two seconds apart, and never searches. Searching Google Patents from a program means a paid intermediary (serpapi) or Google’s own BigQuery patents dataset.

Services looked at and not built

Install

Python 3.11+. Not on PyPI; install from this repository:

pip install "patent-fetch @ git+https://github.com/polarizetech/patent-fetch"
pip install "patent-fetch[s3] @ git+https://github.com/polarizetech/patent-fetch"   # plus the S3-compatible store (boto3)
# or, from a checkout:
uv sync && uv run patent-fetch --help

The core depends only on the standard library. boto3 is needed only for the S3 store.

Configure

Everything is configured with environment variables; none is required. patent-fetch also reads KEY=value lines from *.env files in ~/.config/patent-fetch/ (or $PATENT_FETCH_ENV_DIR), taking only the variables below and never overriding the environment. Keep those files out of repositories (chmod 600).

Everything is stored under the prefix patents/, so one bucket can hold this library beside another that keeps to its own prefix.

Numbers

A patent is named by its publication number, in any common spelling: US7861317B2, US 7,861,317, EP 2264377 A2, WO 2012/016009, US 2012/0029596 A1, or a Google Patents URL. It is stored under country code plus number without the kind code (US7861317). A US application number (US 10/962,874) is refused, as is a number with no country code.

Python

from patent_fetch import Library
from patent_fetch.config import load_env

load_env()  # keys from ~/.config/patent-fetch/*.env, if you keep them there
lib = Library.default()

res = lib.search("cochlear implant electrode array", since=2015, countries=["US", "EP", "WO"])
for report in res["providers"]:  # read these before reading "no hits"
    print(report["provider"], report["status"])
for hit in res["hits"]:
    print(hit["id"], hit["year"], hit["title"], hit["found_by"])

rec = lib.fetch("US7861317B2")  # the library first, then the providers
print(lib.claims(rec["id"]))
print(lib.record(rec["id"])["provenance"])  # which provider, when, every route tried

More in INTEGRATION.md.

Development

uv sync
uv run pytest                     # offline: no credentials, no network
uv run ruff check . && uv run ruff format --check .
uv run pyright

The suite never opens a socket: HTTP is a fake that serves tests/fixtures/. CLAUDE.md records what was measured against each live service and how to add a provider.

Licence

MIT. See LICENSE.