How we check

Every fact on CookiesForAI comes from a named source, such as the publisher's GitHub repository, its docs, the MCP Registry, an App Store or Google Play page, the maker's website or the maker's own Show HN post, and carries the date we read it. A page's "Facts re-read" date changes only when a script has read its facts again, never when we only edit the wording.

What each fact means on CookiesForAI

Fact Where we read it
License The skill folder's own LICENSE file or the license: line in its SKILL.md; if neither exists, the repository license GitHub reports. When they disagree, the page says so.
Installs into these agents We list the skill with the open skills CLI (npx skills add <repo> --list, version 1.7.0 on ). That CLI installs a skill into 78 agents, and our table shows the folder each agent reads.
Works with Only the agents and apps the publisher's own README or docs name.
Install and connect commands Copied from the publisher's README or docs, never written by us.
GitHub stars, last change The GitHub API on the date shown. "Skill last changed" is the last commit that touched the skill's own folder.
Free tier and cost The maker's own pricing page, or its license and docs when there is nothing to pay, on the date shown. We say whether there is a free tier and who bills what, link the pricing page when there is one, and record dated plan prices for some coding tools. We hide prices if our weekly check cannot confirm them.
Official The company whose product it is publishes it from its own account.
Official MCP Registry The server's name in registry.modelcontextprotocol.io, searched on the date shown.

The "Facts re-read" date covers what our script reads again: GitHub stars, last push and archived flag, a skill's last change, version and license file, and the MCP Registry listing. It also scans each pricing page we link for text aimed at AI assistants. Supported agents, install commands and requirements are read from the publisher's README or docs when we write or edit the page; the script does not re-read them.

Which sources CookiesForAI lists

We list a skill's GitHub repository when one of four rules holds:

  1. Official: the company whose product it is publishes it from its own GitHub account.
  2. Widely used: the repository has more than 5,000 GitHub stars.
  3. Named by AI engines: AI assistants from at least three different companies named the repository by its owner or its address in most of their answers to the same question, in at least two answers each, within the last 90 days. We count only answers given in temporary or private chats, or in scripted runs that keep no history and load no user settings, and never an answer that cites CookiesForAI. Several products from one company count as one company. A repository listed this way must also have at least 1,000 stars, an open source license in its own files, a push in the last year, not be archived, and pass our check for text aimed at AI assistants. We ask the question again in our regular checks after launch; a listing comes down when the latest checks stop naming it, and at the latest 90 days after the last run that did.
  4. Trending agent skills: an agent skill is on that day's trending list, its row passed the list's checks, and the repository has at least 1,000 GitHub stars. GitHub and the repository's own LICENSE or SKILL.md must each name an open source license. The repository is not archived, was pushed in the last year, is not listed here yet, and passes our check for text aimed at AI assistants in its README and in the SKILL.md of each skill that gets a page, and the open skills CLI lists that skill. We add at most two pages a day: repositories that gained the most stars go first, and their skills in the order the skills CLI lists them. An AI model writes each page from the README and SKILL.md, code checks every field and that the page copies no line of prose and no ten words in a row from those texts outside quotation marks, and a second AI model checks the page against the same README and SKILL.md text. Our editor approves the new pages with that day's list in one approval before anything goes live. The listing stays after the repository leaves the trending list.

The rules are about where a skill comes from and how many people use it, not a review of its quality.

Which apps and brands CookiesForAI lists

Our Apps section lists products by independent makers: phone apps, web apps, games and small brands with their own shop. We add each one ourselves from public pages; no maker pays to be listed and we do not ask makers first. We list a product only when all of these hold on the day we check it:

  1. Live and released: the website answers on the product's own domain, or the app is on its store page, and it is not waitlist-only.
  2. Recent: the store shows an update in the last 18 months, or the maker's site shows dated activity in the last 12 months.
  3. A named maker: a person or company is shown on the site, the store page or a legal notice. We show only the maker's public name, never an address, email or phone number.
  4. Independent: one public page shows that the App Store seller is a person, or that the maker's own page names 10 or fewer people, or that a legal notice names a sole trader or small company and the About page describes a small team. We also search for funding or a large owner, and we do not list a product when we find one. These thresholds are our own choice, not an industry standard.
  5. Its own product: not a template, a copy of another app, or a thin layer over a general AI chat. It keeps the user's own data, works offline, uses a device feature, or does a specific job a chat assistant does not do in one prompt.
  6. Not excluded: no adult content, gambling, weapons, drugs, crypto tokens or investing, medical diagnosis or treatment, AI therapy, apps that hold or move money or give investment advice, services, courses, newsletters or lists.

Every fact on a product page names the page it comes from, and the page links those pages. We describe what a product does, never what it achieves: we do not repeat claims about health effects, sustainability or quality. "AI inside" and "Built with AI" appear only when the maker says so in its own words on its website, store page or Show HN post, and the product page links that source. A category gets its own page once it holds eight products.

Lists are in alphabetical order. Paying never changes the order or what we write. If you make a product we list, write to hello@cookiesforai.app to correct or remove its page.

MCP numbers

We check whether each listed MCP server answers or needs a key, how many tools it lists, and how many tokens its tool list takes: the name, description and input schema of every tool, as the server sends them, written out as JSON. Assistants that load every tool send this list to the model with each question; some load tools only when a task needs them. For each server with a safe probe, we also make one fixed read-only call and count the tokens in its answer.

The check runs on a throwaway GitHub machine. Program servers run in sealed containers. A full check runs every Monday, and a smaller alive check runs daily. We never use a key or sign in, so a server that requires either has no tool numbers, and a server that lists more tools once you sign in shows only what it listed without. A server that needs a desktop app, a database or a key to work is not run at all. A server page shows the numbers only when we measured them. The chart on a server page sets its numbers next to a typical server from the same run: the median first question (tool list plus one answer) of the servers where we measured both, or, for a server without a one-answer number, the median tool list of the servers that listed their tools.

The counts use OpenAI's o200k_base tokenizer through js-tiktoken 1.0.21. Other model families count the same text somewhat differently. We measure the servers ourselves and never take numbers from another measuring site.

Trending AI tools is a daily, dated measurement. The job runs each evening from about 19:30 Berlin, reading GitHub and Show HN; it can start later when GitHub is late.

GitHub repositories must be created in the last 90 days, have at least 300 stars, gain at least 200 stars over a 7-to-9-day window, or since the repository was created when it is younger than 7 days, and come through 17 AI topics. Ordered by stars gained over the dates shown, each row gives its dates, gain, total stars, the date we read them and any license. For every GitHub and Show HN row, an AI model writes one sentence from the README, and a second AI model checks it against a passage quoted from the README. The top three appear at home.

Show HN follows as a separate group. Posts need a GitHub repository link and at least 50 points within 48 hours. They are ordered by points, highest first.

An AI model asks from the README whether the tool can be installed, opened or run today. We do not run tools. Only runnable apps, command-line tools, coding tools, browser extensions, agent skills, plugins, agents, agent frameworks and MCP servers qualify, not model weights, datasets, papers, courses, link lists, prompt collections or templates with nothing to run.

We exclude the topics our apps rule 6 above excludes and tools whose name or description offers free or unlimited use of paid AI models. Checks may also hold repositories for stars out of line with forks and watchers or READMEs addressing AI assistants. Held tools are left off unnamed. A tool may be missing because a check did not finish that day. Being left off says nothing about a project.

Every day's list is fresh: at most 20 GitHub and 10 Show HN rows. It is not a recommendation or a security check. Some agent skills on the list also get a page in our directory under rule 4. Read a tool's source before installing it. Nobody can pay to appear or rank higher. Nothing goes live until our editor approves each day's list.

What CookiesForAI does not claim

A listing is not a security audit. We do not run listed skills, servers or tools, and we do not grade code safety. A skill can run scripts on your computer: read its source before you install it.

How CookiesForAI writes and ranks

AI models help with research, drafting and checking: they draft descriptions, titles and articles; write the daily news cards and the trending list's one-line descriptions, each checked by a second model against a quoted passage of its source; give a first reading of the lines our pattern scan flags as aimed at AI assistants; and check facts in our weekly re-read. Tests stop a build when a page's license or date disagrees with what the scripts read, or when a page shows a price without the date we checked it. Each line the scan flags is read and the verdict is stored with the item, with who read it; a verdict from an AI model is marked as waiting for a person. We never copy a skill's or vendor's text, and a test fails when our copy repeats a line or a ten-word run of a SKILL.md. In our guides, a "best for" verdict always names the job it is best for and the facts behind it. We never rank a product higher because its maker asked or paid.

How often CookiesForAI re-checks

A job re-reads the directory facts every Monday, and nothing goes live until our editor approves the result. The sources for the trending list are read daily. When a source disappears or a repository is archived, the page says so with the date.