iporigin

Tell a datacenter IP from a home connection — offline, with no API key.

>>> import iporigin
>>> iporigin.classify("52.95.110.1")
Origin(ip='52.95.110.1', kind='hosting', provider='Amazon AWS', source='local')
>>> iporigin.is_datacenter("8.8.8.8")
True
>>> iporigin.is_datacenter("127.0.0.1")
False

No network calls. No signup. No runtime dependencies. The answer comes from a bundled table of 38,000 ranges covering 48 hosting providers, CDNs, consumer VPNs, Tor and declared crawlers.

Install

pip install iporigin

Why

"Is this visitor on a server or on a home connection?" comes up constantly — scoring signups, filtering scrapers out of analytics, deciding whether an abuse report is worth acting on, flagging logins from hosting ranges. The usual answers are a paid API or a hand-maintained list of CIDRs that goes stale in a month.

This is a third option: the providers publish their own ranges, so the list can be compiled from source and rebuilt on a schedule. You get a local lookup in microseconds, and you can verify every byte of the data by re-running the build script.

Use

Classify one address

import iporigin

origin = iporigin.classify("140.82.121.4")
origin.kind          # 'hosting'
origin.provider      # 'GitHub'
origin.is_datacenter # True

kind is one of:

Two sets are exported for the common decisions:

origin.kind in iporigin.DATACENTER_KINDS   # a machine, not a home line
origin.kind in iporigin.ANONYMIZER_KINDS   # vpn or tor

They are deliberately different questions. Someone arriving through Mullvad is a person at home, but the address they arrive from is a server — so it is in both sets, and you may well want to rate-limit one and refuse the other.

unknown means absence of evidence. It is not a positive claim that the address is residential, and the difference matters if you are going to block someone over it.

Scan a lot of them

for origin in iporigin.classify_many(ip_list):
    if origin.is_datacenter:
        print(origin.ip, origin.provider)

classify_many loads the table once for the whole batch.

From the shell

$ iporigin 8.8.8.8 140.82.121.4 192.168.1.1
8.8.8.8                                  hosting   Google
140.82.121.4                             hosting   GitHub
192.168.1.1                              reserved

$ cut -d' ' -f1 access.log | iporigin --datacenter-only --json
{"ip": "34.82.1.5", "kind": "hosting", "provider": "Google Cloud", ...}

iporigin --info prints what is in the bundled dataset and when it was built.

Going beyond the bundled table

The table covers the VPN providers whose exits are publicly tracked — Mullvad, ProtonVPN, Apple Private Relay. Most VPN companies are not in that set. When you need broader coverage, iporigin.online asks the free Unblock Master IP API, which does its own detection:

from iporigin.online import classify_online

classify_online("203.0.113.10")   # may return kind='vpn'

It falls back to the offline answer if the request fails, so it is safe in a request path. Nothing else in the library touches the network — you have to import this module on purpose. No key required.

What is in the dataset

48 providers, in three tiers.

Published by the provider. The authoritative tier — each of these is the company's own feed, fetched at build time:

Community-maintained lists. Some providers publish nothing machine-readable — Hetzner and OVH being the two that matter most, since a large share of abusive traffic comes from them. Consumer VPN exits, Tor and crawler ranges have the same problem for a different reason: nobody with the data has an interest in publishing it. Those come from two community repos, and are second-hand by definition:

• rezmoss/cloud-provider-ip-addresses (CC0)
• lord-alfred/ipranges (CC0)

Covering Hetzner, OVHcloud, Scaleway, Alibaba Cloud, Leaseweb, UpCloud, IBM Cloud, Huawei Cloud, Tencent Cloud, Rackspace, Akamai, Gcore, Mullvad, ProtonVPN, Apple Private Relay, Tor, and ten declared crawlers. AI-company outbound ranges (OpenAI, Perplexity AI, DuckAssistBot, Apple Intelligence Proxy, Meta) come from these same two repositories. Most of them overlap exactly with an existing bot label, so they contribute label diversity rather than new bot coverage — only Meta adds distinct segments.

Both are CC0, a public-domain dedication, so they carry no conditions.

Used with the maintainer's permission. Two further repositories publish no licence file, which normally rules them out. They are included because permission was obtained from each maintainer directly — see NOTICE, which records what was granted and by whom:

• 123jjck/cdn-ip-ranges — Cogent, DataCamp, Contabo, Vercel, CDN77, GleSYS, Scalaxy, GTHost, Melbicom, BuyVM, BunnyCDN
• SecOps-Institute/Akamai-ASN-and-IPs-List — additional Akamai ranges

Only what is additive is taken. jhassine/server-ip-addresses is the best-known list of this kind and is not included: 52,772 prefixes, 227M addresses, and every one of 211,616 sampled addresses was already covered. It is a subset of the provider feeds it was itself built from, and a source that adds nothing is still another endpoint that can break the weekly rebuild.

About 454,000 published prefixes collapse into roughly 38,400 disjoint ranges (about 21,900 IPv4 and 16,500 IPv6) covering 287 million IPv4 addresses. The exact figures move every week with the feeds; the build prints them. Rebuild it yourself at any time:

python tools/build_dataset.py

A GitHub Action re-runs that weekly and commits the result when the ranges move, and a second one cuts a patch release on the 6th of each month if the dataset changed since the last release — so pip install --upgrade iporigin is never more than about a month behind the feeds, and a quiet month produces no release rather than a version whose only content is a new number. The build refuses to replace the committed dataset if it shrinks by more than 20% — a feed that starts answering with an empty body looks exactly like a provider giving up its address space, and nothing else would catch it. Pass --allow-shrink when the drop is genuine.

Known gaps

Being explicit about these is more useful than pretending they are not there:

• Most consumer VPNs. Only the ones whose exits are publicly tracked are in the table. Use iporigin.online for the rest.
• Some provider-owned addresses sit outside the ranges the provider publishes. 1.1.1.1 is Cloudflare's resolver but is not in Cloudflare's published edge list, so it comes back unknown.
• Second-hand data is second-hand. Tiers 2 and 3 are as good as those repos are, and they are not the provider speaking.
• The data is as accurate as the feeds. A range reassigned yesterday is wrong until the next rebuild.

How the lookup works

Ranges are stored as inclusive integer start/end pairs in sorted, disjoint order, so a lookup is one bisect plus one comparison.

Making them disjoint is the part that matters. Feeds overlap each other — GitHub runs on Azure and AWS, so its prefixes sit inside theirs. A bisect inspects exactly one candidate, and with overlapping ranges that candidate can be a narrow range that ends before the address while a wider range still contains it, which returns unknown for an address plainly in the table. The build script therefore sweeps the ranges into a disjoint partition, and where they overlap the narrowest one wins — GitHub inside Azure answers GitHub, which is the more specific truth.

The table loads lazily on the first lookup, so importing the library and never calling it costs nothing.

Compatibility

Python 3.8+. No dependencies.

License

The code is MIT. The bundled dataset comes from third parties; every source, its licence, and — where there is none — the permission it is used under are recorded in NOTICE.

Built by Yuix Networks, who also run Unblock Master and its free IP lookup API.