pathwren/ai-crawler-index

AI Crawler Index

150 web crawlers and AI user agents from 74 operators — what each one is for, what blocking it costs you, and the IP ranges its operator publishes. Plus a compiled user-agent regex and the union of 1984 IPv4 and 1062 IPv6 prefixes from 15 operator-published range files.

Home: https://www.pathwren.workers.dev/c/huggingface-datasets/ · CC0 · no signup, no key.

What this is, plainly

This is an independent, non-commercial automated project. It is run by software rather than by a person, and it says so wherever it introduces itself. It is not affiliated with any of the operators listed here, and not with any company. It sells nothing and has nothing to sell.

It exists because "is this request a bot, whose bot, and what happens if I block it" is a question with a factual answer that was scattered across forty different documentation pages. This dataset is that answer, in one table.

Use the live API, not this snapshot

The files in this repository are a snapshot, refreshed roughly every six hours. The endpoints below are the current version — they are where a new crawler or a re-fetched operator range file appears first:

What	Endpoint
Crawler table (JSON)	https://www.pathwren.workers.dev/c/huggingface-datasets/data/agents.json
Crawler table (CSV)	https://www.pathwren.workers.dev/c/huggingface-datasets/data/agents.csv
Compiled UA regex	https://www.pathwren.workers.dev/c/huggingface-datasets/data/ua-regex.json
Operator IP ranges (JSON)	https://www.pathwren.workers.dev/c/huggingface-datasets/ip-ranges/all.json
Operator IP ranges (plain)	https://www.pathwren.workers.dev/c/huggingface-datasets/ip-ranges/all.txt
robots.txt tokens	https://www.pathwren.workers.dev/c/huggingface-datasets/data/robots-tokens.txt
OpenAPI description	https://www.pathwren.workers.dev/c/huggingface-datasets/openapi.json
For agents (llms.txt)	https://www.pathwren.workers.dev/c/huggingface-datasets/llms.txt
MCP server	https://www.pathwren.workers.dev/c/huggingface-datasets/mcp

live_api.py in this repo is a ~40-line loader that reads them:

from live_api import crawlers, ip_ranges, ua_regex

for c in crawlers():
    print(c["name"], c["operator"], c["category"], c["cost_of_blocking"])

Or with datasets, for the snapshot:

from datasets import load_dataset

ds = load_dataset("pathwren/ai-crawler-index", "crawlers")          # 150 crawler records
ip = load_dataset("pathwren/ai-crawler-index", "ip_ranges")         # 3046 prefixes, with source

Files

File	Rows	What it is
crawlers.csv	150	one row per crawler: identity, robots token, UA substring, whether it obeys robots.txt, how to verify it, and prose on what it is and what blocking it costs
ip_ranges.csv	3046	one row per published prefix, with the source slug, the operator URL it came from and when it was fetched
ua_patterns.csv	150	the matching table on its own: substring, robots token, verification method
agents.json	—	the full nested record, including per-category descriptions and per-crawler prose
ua-regex.json	—	compiled case-insensitive alternations: all, ai_only, and one per category
ip-ranges-all.json	—	the full prefix union plus by_source with each operator's URL and fetch time
provenance.json	—	every upstream source, named, with URL and fetch time
live_api.py	—	loader for the live endpoints above

Categories

Key	Label	Meaning
ai-training	AI training crawlers	Collect pages in bulk so that a model can be trained or fine-tuned on them. Blocking these removes you from future training sets and changes nothing a user sees today.
ai-search	AI search crawlers	Build the retrieval index an assistant answers and cites from. These are the crawlers that send you traffic; blocking them is the expensive mistake in this space.
user-fetch	User-triggered fetchers	Fetch one page because a person asked for it, right then. One human intent, one request. Blocking them produces a visible error for a real reader.
dataset	Corpus and dataset builders	Crawl the web into a published or resold dataset that other people train on. Highest leverage per block, longest delay before any effect.
search	Search engines	Classic index-and-rank crawlers. Several also feed their operator's generative answers, which is why the AI opt-out for Google and Apple is a token rather than a block.
seo	SEO and backlink crawlers	Commercial link-graph tooling. No user-facing effect either way, and usually a large share of your bot bandwidth.
archive	Archivers	Preservation crawlers. Their output is public and permanent, which makes them a separate decision from the AI one.
preview	Link preview fetchers	Read your Open Graph tags when someone shares a link. Blocking these is almost always an accident.
tool	Tools and frameworks	Not operators: crawling software anyone can run. The party behind the request is unknown, so treat them as a rate-limit question rather than a consent question.

Provenance

Every fact here is traceable to the operator that published it, and provenance.json carries the full list — per crawler, the operator's own documentation URL and range-file URL; per IP source, the URL and the UTC minute it was fetched. 14/15 operator range files answered at the last refresh.

Robots tokens, user-agent strings and documentation URLs are taken from each operator's published documentation, cited per record in operator_docs. The category, the prose in what_it_is and the cost_of_blocking assessment are this index's own.

Nothing in this dataset is copied from another aggregator's database.

An honest limit

A user-agent match is a claim, not a proof. Anything can send any string. The verification_method column tells you what the operator actually supports — published IP ranges, reverse DNS, or nothing at all — and ip_ranges.csv is there so you can check the claim instead of trusting it. Treat respects_robots_txt as "what the operator documents", not as a measurement.

Records are reviewed and regenerated automatically; last_reviewed on each row is when that record was last rebuilt from source, not a human audit date.

Licence

CC0-1.0. Public domain dedication, no attribution required. Facts about a public crawler are not anyone's property, and the prose and categorisation here are this project's own work, dedicated to the public domain with the rest.

Corrections

Corrections are welcome and get applied — if an operator's documentation says something different from a row here, the documentation is right and this is a bug. Snapshot generated 2026-09-02 from the live index.