stack.tools

Methodology

The rule behind everything on this site: a fact without evidence does not ship. Here is exactly what that means.

What counts as a fact

Every entry is a structured claim — “company X runs tool Y in production” — that carries all of the following, or it is not published:

company + tool + vendor
canonical product name, vendor listed separately ("Claude" / "Anthropic")
layer
exactly one layer from the public taxonomy
usage
one line on what the tool actually does for them
evidence
at least one receipt: URL, source type, publication date, verbatim quote
confidence
high or medium — nothing weaker is published
status
current, or historical when they have moved off it
last verified
the date the source URL was last confirmed live and still supporting the claim

Quotes are verbatim excerpts of the source, capped at about twenty words — never paraphrased, never reconstructed. If the source does not literally say it, we do not quote it.

Where facts come from

AI stacks are backend-invisible: no crawler can inspect which model or vector database a product calls. So there is no BuiltWith-style detection here — only what companies state in public, ranked by evidentiary value:

  1. 1

    Job postings

    "Experience with LangChain / Pinecone required" is dated, first-party proof that a company runs a tool.

  2. 2

    Subprocessor lists & trust centers

    Legally maintained vendor lists (GDPR). When a company names its AI vendors here, that is as primary as it gets.

  3. 3

    Engineering blogs

    "How we built our RAG pipeline" posts, in the company's own words.

  4. 4

    Vendor case-study pages

    Customer stories published by the vendor, inverted into company records. First-party from the vendor side; we date them.

  5. 5

    Podcasts & conference talks

    Engineers describing their stack on the record (transcribed). Quoted verbatim from the transcript.

  6. 6

    Public GitHub repos

    Dependency files and infra code in repos the company publishes.

  7. 7

    Community threads (HN, Reddit)

    Treated as LEADS ONLY — never published from a thread alone; every lead is verified against a source above.

  8. 8

    Self-submission

    A bonus signal, never a substitute for public evidence.

Evidence types

Every receipt is tagged with its source type, shown as a badge next to the link:

Engineering blog Job posting Subprocessor list Vendor case study Podcast Conference talk GitHub Docs Press Social

Confidence levels

High confidence

A primary source, unambiguous: the company (or its vendor, on the record) states it directly, and the statement is recent enough to trust.

Medium confidence

A secondary or dated source: solid evidence, but second-hand or old enough that the stack may have moved on. Shown as-is, labeled as-is.

Below medium

Dropped. Not hedged, not footnoted — not published. We would rather list 8 well-evidenced facts than 30 weak ones.

Freshness & historical facts

Stack directories die of staleness, so freshness is treated as part of the product. Every fact displays a last verified date — when its source URL was last confirmed live and still supporting the claim — and sources are re-checked on a schedule. When evidence shows a company moved off a tool, the fact is not deleted: it is marked historical and stays visible, because migrations are often the most interesting data point on the page.

The dataset

Everything on this site — every company, fact, and receipt — is available as a single machine-readable JSON export, regenerated on every build. Licensed CC BY 4.0 with attribution to stack.tools.

Download the dataset (stacks.json)

Corrections

If you work at a listed company and a fact is wrong or outdated, tell us — we will re-verify against the sources and fix it. Facts that fail re-verification are corrected, marked historical, or removed.

Email a correction

What this site does not know

Honest limits: coverage is only as complete as what companies say in public — absence of a tool here is not evidence a company does not use it. Facts can lag reality between re-verification passes. And this site is maintained by an autonomous agent, curated and reviewed by Denis Lunev — the verification gate exists precisely because extraction can be wrong. How that pipeline works is documented on the operations page.