How Kapyn is made

Kapyn claims two things on every page: that it is curated, and that it is kept current. This page is the evidence, including the parts that are automated and the parts we deliberately do not claim.

The catalog: what is hand-written

The spine of the Radar is written by hand, one entry at a time. Right now that is 144 tools, 73 MCP servers (53 first-party, 20 community) and 48 skills. Every one has a human-written line saying what it is for. None of them is generated.

These numbers are small on purpose. A bounded list someone has actually read is the product; an exhaustive index is what we are trying not to be. It also means the catalog has gaps, and a tool missing from it is not a judgement on that tool.

The catalog: what is discovered automatically

Alongside the hand-written spine, a daily job surfaces what is newly rising: GitHub repositories, Product Hunt launches and the official MCP registry. These are held to a separate standard and labelled differently from curated entries. Registry-discovered MCP servers currently need at least 10 stars to appear at all, and a curated entry always wins over a discovered duplicate of itself.

The daily brief: how a story gets in

1. A scheduled job reads more than thirty publisher feeds, from the labs themselves (OpenAI, Google DeepMind, Microsoft Research, Hugging Face, Mistral) through to the technology press, plus Product Hunt.
2. Duplicates are dropped on the source URL, so the same story cannot appear twice.
3. Each item is classified into a category and summarised by a language model, working from the publisher's own text. This step is automated. Summaries are not written by a person.
4. An automated quality gate then rejects the output: anything the model flagged as low-signal or off-topic, anything that leaked the prompt, and raw changelog boilerplate. Rejected items are not published at all rather than published badly.
5. What survives is stored with its source name and a link out. The live feed shows the last 7 days; everything older stays permanently addressable at its own story URL.

The honest limitation: a model summarising a headline can be subtly wrong in a way the gate will not catch. That is why the source link is on every card and why we would rather you click it than trust us.

What “trending” and “momentum” mean

Neither is based on Kapyn's own traffic. We do not rank by what our readers click, because at our size that would mostly measure noise.

• Trending stories are scored by how many distinct outlets covered the same lead subject, multiplied by a recency decay. Broad coverage of a fresh subject outranks a single outlet.
• Momentum on Pulse is how recently a tool was seen rising, combined with its repository or launch score. It favours the new and fast-moving over the merely large, which is the point, and it also means an established tool can be absent from Pulse while remaining excellent.

Keeping the model comparisons current

The comparison pages track 19 model families. They were last reviewed on 2026-10-21, which is printed on the page itself rather than hidden.

Two deliberate choices make that claim survivable. Models are listed by family, with the current release as a separate field, so a point upgrade is a one-line edit instead of a rewrite. And prices are shown as qualitative tiers with a link to the provider's live pricing page, because a hard per-token figure in our copy goes stale silently and we would rather send you to the source of truth.

A check runs in CI and fails the build when the model data has not been reviewed recently, when a family has no current release recorded, or when a retired model name survives in a live recommendation. What it cannot do is notice that a provider shipped something: that still takes a human looking. Treat a review date as “recently checked”, never as a guarantee of being current.

The explainers

The concept explainers are generated from stories Kapyn has already ingested, and each one carries the number of sources it was built from and the date it was last updated. They are machine-drafted from sourced reporting, which is stated on the page, not disguised.

What we do not claim

• Not that the catalog is complete. It is bounded by design and it has gaps.
• Not that we have tested every tool against its alternatives. The alternatives pages list same-category peers from the hand-written catalog; they are a starting point, not a verdict on substitutability.
• Not that summaries are human-written. They are not, and the pipeline above says so.
• Not that everything is current to the day. Freshness is a job we run, with a date you can read, not a property we assert.

More on who maintains Kapyn and what it commits to: About Kapyn.