OKF vs llms.txt vs AGENTS.md vs JSON-LD
These four formats often get lumped together because all of them "let machines say something about content." But they solve different problems in different places. This page sorts them along clear dimensions without crowning a winner – each fits a different job, and most of them combine.
The short version
| Dimension | OKF v0.2 | llms.txt | AGENTS.md | JSON-LD |
|---|---|---|---|---|
| Einheit / Unit | Konzept-Bundle (viele Dateien) / concept bundle | eine Datei je Site / one file per site | eine Datei je Repo/Ordner / one file per repo | Annotation je Seite / per-page annotation |
| Hauptkonsument / Primary consumer | Mensch + RAG-Agent / human + RAG agent | LLM zur Inferenzzeit / LLM at inference | Coding-Agent / coding agent | Suchmaschine, Knowledge Graph / search engine |
| Inhalt / Content | Prosa + Metadaten / prose + metadata | kuratierter Link-Index / curated link index | Anweisungen / instructions | typisierte Entitäten / typed entities |
| Provenance/Trust | ja (generated, verified, status) | nein / no | nein / no | begrenzt (Autor, Datum) / limited |
| Versionierbar in Git / Git-diffable | ja / yes | ja / yes | ja / yes | im Seitenquelltext / in page source |
| Herkunft / Origin | offenes Format, Google-Basis + Profil | Answer.AI (2024) | Linux Foundation / AAIF | W3C / schema.org |
OKF – the knowledge bundle
OKF describes knowledge as a collection of individual concepts: one Markdown file with a YAML header each, together a bundle. The focus is on the content itself and its dependability – with the v0.2 layer of provenance, review, and freshness. A bundle is simultaneously human-readable, git-diffable, and built for AI retrieval. The strength: granularity and trust. The limit: it is not a site signpost or an agent manual, but the body of knowledge beneath.
llms.txt – the signpost at the site root
llms.txt, proposed in 2024 by Jeremy Howard (Answer.AI), is a single Markdown file at /llms.txt that gives a language model a curated overview of a website at inference time: an H1 with the name, a short blockquote summary, then H2 sections with link lists to the important pages (often mirrored as .md). The purpose is findability, not storage: the model should quickly locate where the substance lies. llms.txt complements sitemap and robots.txt rather than replacing them.
llms.txt and OKF do not clash. An OKF bundle provides the concepts; an llms.txt can point at them. This project's pipeline produces both – the concepts plus an llms.txt as a crawler signpost.
AGENTS.md – the manual for coding agents
AGENTS.md is an open format that tells a coding agent how to work in a repository: build commands, test commands, code conventions, security notes. A "README for agents," now used by over 60,000 open-source projects and stewarded under the Linux Foundation. It describes instructions, not a body of knowledge. That puts it orthogonal to OKF: one tells an agent how to work a project, the other supplies what it needs to know.
JSON-LD – the typed annotation for search engines
JSON-LD per schema.org embeds structured data into a web page: typed entities and their relationships, in a <script type="application/ld+json"> block. Search engines and classic knowledge graphs read facts about products, people, organizations, and articles from it. The strength lies in the machine precision of typed fields – a price is a price, an author is an author. The limit is prose: JSON-LD annotates content, it does not carry it. A prose body of knowledge with derivations and examples fits poorly into typed fields.
Along the deciding dimensions
Granularity. JSON-LD works at field level, OKF at concept level, AGENTS.md and llms.txt at file or site level. The finer the unit, the more targeted the retrieval – and the more maintenance.
Provenance and trust. Here OKF v0.2 stands alone: generated, verified, status, stale_after, and sources make origin and freshness machine-readable. JSON-LD can carry author and date but no review and lifecycle. llms.txt and AGENTS.md do not address the topic – they are not meant to.
Portability. OKF, llms.txt, and AGENTS.md are plain-text Markdown and live in git. JSON-LD lives in the page source and is bound to the page it sits in. For a versioned, shareable body of knowledge, the plain-text approach has the edge.
When which
A realistic setup uses several at once. JSON-LD annotates the public website for search engines. llms.txt orients language models at the site root. AGENTS.md steers the agents working on the code. And OKF carries the actual body of knowledge a RAG system queries – with the provenance the other three do not provide. The question is rarely "which format" but "which layer" – and OKF occupies the knowledge layer.
Continue
How retrieval over an OKF bundle is measured is on the RAG benchmarks page. The format itself is explained in the OKF v0.2 specification.
Quellen: llms.txt (Answer.AI, 2024), AGENTS.md (Linux Foundation), schema.org / JSON-LD, OKF SPEC.md. Abgerufen am 2. August 2026. / Sources retrieved 2 August 2026.