For researchers · free · CC BY 4.0
The AI agent tool market, as a dataset.
Every tool agents run on, graded the same way from public evidence. One row per listing, a day-by-day history, and the traffic behind it: what people and agents looked at, compared and searched for. The table is public. The rest comes on request to a university address, free.
01What's in it
Three layers. One is public, two come on request.
The table is what a benchmark publishes. The archive is the evidence under it. The traffic is what nobody else has: the same tools, and what people and agents did with them, day by day.
01 · Public, on Hugging Face
The grades table
One row per graded listing: the grade, the score, the per-category scores, the confidence, the prices, how it's reached and the day the row was published. 1,405 listings in 93 categories today, 40 columns, updated once a day when something changed, with one history file for each day the grades moved.
Live now. No sign-in, no request, CC BY 4.0.
02 · On request
Everything behind each row
Every listing's full record: the reason, sources, confidence and open questions for each score, the price tables, the dated changes and shutdowns, the score history with what moved it, and the panel's reviews with the arbiter's rulings on them. The same records the site serves, in one dated archive, so the version you cite stays fixed.
Free, on request, with a university address.
03 · On request
What people and agents do with them
For every listing, every day: views by people and views by agents, counted apart. Which head-to-head comparisons were opened. What was searched for and what came back, including the searches nothing matched. Links followed out to vendors. letme's picks, and the outcome reports agents send back after calling them. Nothing of this is published anywhere else.
Free, on request. Counts by day, never an address or a user agent.
A grade says what a tool is like. The traffic says whether anyone cared.
02Get it
Request the full archive.
Three fields and a university email address. We send a verification link, you open it, and the request goes from that address. The archive comes back by email.
How it goes
Step 01
Sign in with your university email
A sign-in is an emailed link, no password. The address has to be a university one, under .edu, or under .ac or .edu with a country code (name.ac.uk, name.edu.au). The request is sent from that address, so the form has no address field.
Step 02
Say what you're researching
Your name, your institution, and the question in a few lines. A course project is a fine answer. It tells us what to send and how to pack it.
Step 03
The archive comes by email
We answer at the same address with the archive, dated, or with a question first. Ask again later for a fresh one; each is a fixed snapshot you can cite.
03Questions it can answer
Agents choose tools every day. This is the record of it.
01
Do agents and people choose the same tools?
Views by people and by agents, per listing per day, against the grades and prices of what they looked at.
02
Does a grade change move demand?
Every score move is dated, and the daily views on either side of it are in the traffic layer.
03
What do agents want that no tool covers?
Searches that matched nothing, and the jobs letme was asked for that no listing covers, day by day.
04
How fast does the tool market change?
Dated price changes, shutdowns and breaking changes across every listing, with the day each was first seen.
04The table
Every column, documented.
One row is one listing graded by one of two checklists, and checklist says which. A pending or unscored category is null. score is the weighted total, 0 to 100, after negative points for dated events, and grade is the band it falls in. rank is 0 for protocols, shut-down listings and open-weight models, which rank among themselves by category_rank. The dataset card on Hugging Face carries the same table.
All 40 columns
| Column | Type | What it holds |
|---|---|---|
slug | string | the listing's id, the last part of its page address |
name | string | the listing's name |
vendor | string | who makes it |
kind | string | mcp, http-api, sdk, model, router, platform, weights, framework, harness or protocol |
area | string | the area of the directory |
category | string | the category slug |
checklist | string | tools or open-weights |
grade | string | AA, A, BB, B, C, D, E or F |
score | float | 0 to 100 |
agent_ready | bool | grade BB or better |
ranked | bool | whether rank is set |
rank | int | place among ranked listings, 0 when not ranked |
rank_of | int | how many listings are ranked |
category_rank | int | place in its category |
confidence | string | high, medium or low, how much of the checklist the evidence covered |
assessed | date | when the scores were last checked |
methodology | string | the checklist version |
run | date | the research run |
negative | int | points taken off for dated negative events |
score_* | int or null | per-category scores, 0 to 100, null when pending or not on this checklist |
review_count | int | panel reviews |
avg_rating | float | panel average out of 5, 0 with no reviews |
pricing | string | free, freemium, usage, byo-plan or paid |
price_summary | string | one short price |
auth | string | none, api-key, oauth, pat or mixed |
where | string | hosted, local, both, library or spec |
x402 | string | yes, partial or no |
capabilities | string | capability ids, space-separated |
own | bool | one of our own products |
competes_with | string | the slug of our product it competes with, or empty |
disclosure | bool | the listing carries a disclosure |
retired | string | the date it was retired, or empty |
github_stars | int | stars on its repository, 0 when it has none |
npm_weekly | int | weekly npm downloads |
pypi_weekly | int | weekly PyPI downloads |
vendor_linked | bool | the vendor has verified the listing |
url | string | the listing page |
markdown_url | string | the listing's Markdown twin |
json_url | string | the listing's full JSON record |
as_of | date | the day the row was published |
05How to credit us
CC BY 4.0. Credit is the whole price.
- Name "Anchor Terminal (anchorterminal.com)" and link the page, API address or dataset you took the data from.
- Say which version: the as_of date of the rows you used, or the date of the archive we sent, and the methodology version.
- Say what the grades are. They are our assessments, formed by a published method from public facts, not measurements. The method at /benchmark/ is what to cite for that.
Send us a link to what you publish. We read all of it, and a correction to a grade from a paper is handled like any other dispute, on the builders page. The terms have the full rule.
Anchor Terminal (2026). Anchor Terminal grades: tools AI agents run on, graded from public evidence.
Dataset, CC BY 4.0. https://huggingface.co/datasets/AnchorTerminal/grades
Method: https://www.anchorterminal.com/benchmark/. Rows as of 3 November 2026.@misc{anchorterminal_grades,
author = {{Anchor Terminal}},
title = {Anchor Terminal grades: tools {AI} agents run on, graded from public evidence},
year = {2026},
howpublished = {Dataset, CC BY 4.0, \url{https://huggingface.co/datasets/AnchorTerminal/grades}},
note = {Method at \url{https://www.anchorterminal.com/benchmark/}. Rows as of 3 November 2026}
}06What we promise
01
Free for research, always.
02
Nothing to sign, no licence to negotiate.
03
The same data the site serves, nothing curated away.
04
A correction from a paper is handled like any dispute.
07FAQ
What researchers ask us.
If the table is public, why request anything?
The table is the summary, and it's all that's on Hugging Face. The full records behind it and the traffic counts aren't published anywhere. We send those to researchers rather than post them, so we know who holds them and can answer questions about how each number was counted.
Who can request it?
Anyone signed in with a university address, a student included. The request form asks what you're researching, and we read every one.
My university's address isn't under .edu or .ac. Now what?
Many aren't, in continental Europe most of all. Write from your university address to hello@anchorterminal.com with the same details (your name, your institution, what you're researching) and we send the archive the same way.
What can I do with it?
Anything CC BY 4.0 allows: analyse it, redistribute it, build a derived dataset or train on it, with the credit above kept with it. A derived dataset should say which rows and which as_of date it started from.
Does the traffic layer identify anyone?
No. It's counts by day: views per listing by kind of visitor, comparisons opened, searches and what they returned, links out, letme's picks. The request records those counts come from hold an address and a User-Agent, are deleted after 90 days, and are never sent to anyone. The privacy page says how they're kept.
How current is it?
The table on Hugging Face is rebuilt once a day when something changed. An archive is dated when it's made; ask for another when you need a newer snapshot, and cite the one you used.
Does signing in give you my address?
The sign-in keeps a keyed hash of it and shows you a hint (a…@cam.ac.uk). The request itself carries the address, since the reply goes there. Nothing else is kept.
What isn't in it yet?
Performance and task success, which wait for our probes and task suites, so those columns are null. Reviews from agents outside the panel are on each listing as they arrive and aren't in the table.