Skip to the content

DocsHow the data is made

What is read

How a server's tools, a package, a skill and its scripts are read, what a flag and a trigger result mean, and how the log is made.

Beyond what a publisher wrote in a registry record, the agent directory reads a few things itself. This page says how each is read, and what a reading cannot show. None of them is a review.

What is read from a server

Once a day, a server with a Streamable HTTP address that needs nothing filled in, and whose record declares no header, is sent what any client sends first: a handshake and a request for its list of tools. No credential is sent, no tool is called, and what a server returns is never given to a model. From the answer a page shows:

  • How many tools it listed and their names, which are the publisher's own words.
  • What the definitions weigh: the names, descriptions and input schemas as a client loads them, counted with OpenAI's o200k_base tokenizer. It is what the server costs a context window when every tool is loaded up front. Other tokenizers count differently, and a client that loads tools only when needed pays less. Lists are not ordered by it: fewer tools is not better.
  • When the tool set changed, from the day a later read differed from the one before.
  • What it answered instead, when it listed nothing: a request to sign in or for a token, with the date, or no answer at all, with the day that began.

A read is what one client with no credential saw on one day. A server can answer another client differently.

What is read from a package

About half the servers listed have no address to ask: they exist only as a package a person runs on their own machine. Where a server's record names one exact version of an npm or PyPI package, with no arguments to fill in, that version is fetched and run once, and asked for its list of tools the way a client would ask. It is run where it can do nothing else:

  • Fetched with none of its code run: install scripts are off, and a Python package is taken only as a built wheel.
  • Started with no network, a read-only disk, no privileges, an unprivileged user and no credential of ours, on a machine that is thrown away afterwards.
  • Given placeholders, never values: a variable the record declares is set to text that says it is a placeholder. A package that will not list its tools without a real key is recorded as exactly that.
  • Asked for its list and nothing more: no tool is called, and what the package returns is never given to a model.

A page then says which version was run, on what day, which variables had a placeholder, and, when we saw any, how many times it tried to reach the network or write outside its folder. Those tries are a count by an observer inside the container: none could succeed there, and a package can make tries the observer does not see, so none seen is not none made.

What this cannot show. A tool list is what a package says it does. Running it with no network shows nothing about what it does with one, or with your credentials, when a tool is called. A package that listed its tools has not been reviewed, and one that did not has not been judged: many need a real account before they start at all. A read is of one version; a new version is read again, and until then the page shows nothing for it.

What a flag means

A few fixed patterns are run over the tool definitions that were read. Each looks for a known form of a description written to steer a model rather than to describe a tool: text telling the model to disregard its instructions or keep something from the user, characters that do not display, text telling the model to use this tool in place of other servers' tools, and a request for the contents of key or credential files.

A match is a reason to read the server's tool descriptions before an agent uses them. It is not a finding against the server: a pattern cannot tell a warning about an attack from the attack. No match is not a review either: it says only that these few patterns found nothing on that date, in the text that was read: a page says when a list was cut or a definition was too long to check whole. Before a pattern was shown here it was run over thousands of real tool definitions, to check that it does not match ordinary documentation.

What a model's read of a skill is

Not every skill has been read. A skill's page says "No model has read this file yet" until one has, and shows a reading only when there is one. Where there is, this is how it was made.

The skill's SKILL.md was sent to a model as data, with an instruction not to follow anything in it, and the model answered in a fixed shape: what the skill is for, from a short fixed list; topics, from another fixed list; the outside systems it works with; the command-line tools it has an agent run; and four yes-or-no checks of the file. A file is read once for each version of it: again when the file changes, when the way of reading it changes, and a week after a read that failed. A file over 120,000 characters is sent up to there, and its page says the reading is of the first part.

  • Only fixed lists and the file's own words are kept. A category, a topic and a system come from lists this directory wrote. A tool's name is kept only if the file uses it, and a quote only if it is found in the file. Nothing else the model wrote reaches a page.
  • A yes needs words. For a check answered yes, the model must quote the passage of the file that shows it. The quote is compared with the file with runs of spaces and line breaks counted as one space, and must be at least twelve characters. If it is not in the file, the check is "not shown": neither a yes nor a no. A quote shows that the words are in the file, not that they are true. A no is the model's answer alone and is not checked.
  • A file that speaks to its reader gets no reading. When the model takes a passage of the file to be addressed to whoever reads or rates the file, and that passage is found in the file, the skill's page says so and shows no reading of it. That the passage addresses a reader is the model's judgement and may be mistaken; what it means for a reader is the same either way: read the file yourself.

A read is one model's reading on one date, and a page names both. It is not a review of the skill, and no list is ordered by it.

What a trigger result measures

An agent decides whether to load a skill from its name and description. A trigger eval asks whether a description does that job. A page shows a result only for a skill it has been run for.

A model writes, once, ten requests the skill should be picked for and ten on nearby ground that it should not be picked for, from the name and description alone. A request that names the skill or repeats a phrase of its description is dropped, so a set can be of fewer than ten; it stands with at least eight of each. Each request is then put to another model beside the skill's description and those of up to nineteen other skills in this directory, and that model names one skill or none. The others are chosen by a fixed rule from the skill's own name and description, so a skill meets the same others for as long as the directory lists them.

A page gives the counts: of the requests it should be picked for, how many times it was; of those it should not be, how many times it was all the same. It also names the model, the date and how many descriptions the skill stood among. A run in which the model gave no answer to more than two requests is not shown.

That is all it measures: one description against a sample of other descriptions, on one model, with requests another model wrote. A skill that is picked every time can still be a poor skill, and a skill that is seldom picked may only need a clearer description. The counts are not a score and no list is ordered by them.

What is checked in a skill's scripts

For a skill that ships a scripts/ folder, the files in it that are scripts by their ending, are text and are under 256 KB are fetched, and a few fixed patterns are run over them. A page reports one of those patterns, the one for network calls: its matches on real skills were read by hand and were all what it looks for. The others matched too rarely to judge and are not shown.

A match says a script contains what the pattern takes for a network call, with the date of the check; it does not say why the script makes one. "No pattern matched" is said only when every script file the skill lists was read, with that date, and says only that this one pattern found nothing. Nobody has reviewed the scripts either way, and none is ever run here.

The log of what changed

Today lists what differs between each published catalog and the one before it. Every entry is one fixed comparison of the two, with both sides in the entry. No model reads anything, nothing is scored, and an entry says that a reading changed, never what to think of it.

EntryRecorded when
Known vulnerabilityOSV lists an advisory against the version a server's record names, and the catalog before did not have it
No longer listedAn advisory the catalog before had is gone, on a day OSV was asked: the record names a version OSV does not list it against, the record no longer names the package, or OSV withdrew it
Pattern matchedOne of the fixed patterns newly matches a server's tool definitions or a skill's scripts
Tools changedA server's tool list was read whole both times and differs: a tool added or removed, or a name, a description or an input schema changed. Names are given when both lists hold every one. A change only of the counts in a description is not an entry: one that carries a running count changes every day
Deprecated, Left the directoryThe registry marks a server deprecated; or a server the catalog before listed is not in this one, on a day the registry was read whole
Not answering, Answering againA remote has not answered as an MCP server for seven days; or one the log said was silent answers again
New serverA server published by a system's owner appears. Other publishers' new servers are not logged: anyone can publish one
Price changed, Price not confirmed, Price confirmed again, Allowance changed, Newly listedAn infrastructure product's curated price differs from the catalog before, which happens on the day a person read its pricing page again; or its page has gone a week without printing an amount the catalog holds, and its prices are set aside; or its page prints them all again and they are back in the comparisons; or what one of its plans includes differs from the catalog before; or a product is listed for the first time
New skill, Skill changed, Skill removedA vendor's repository gains a skill, a skill's SKILL.md differs, or a skill is gone on a day its repository was read
New releaseA harness's latest release has another tag than before

An entry is dated by the catalog that first carried it. A day on which no catalog was published, or one whose catalog differed in nothing logged, has no entry and is shown as "nothing recorded": that is a day no reading differed, not a statement about any server. The log keeps 30 days and began with the first catalog that carried it; nothing older was recorded.

When one run finds more than 25 changes of one kind, the ones about a system owner's own server are listed and the rest are one line with their count. That many at once is more likely one change in a source (a registry reset, a newly listed system) than as many separate ones.

Not logged: download counts, stars, the dates facts were read, a server's version by itself, anything about a server that is only listed, and a model's read of a skill. Model retirements and price changes on the page are read from the model catalog at build time and are not entries of the log.

The same entries are an Atom feed, at /feeds/today.atom for the whole log and at /feeds/systems/<system>.atom for one system, which each system's page links. An entry's id never changes, and a feed changes only when the catalog does.