Portable grounding for AI

What is OpenMOCA?

OpenMOCA is an open standard for portable grounding for AI. MOCA stands for Modular Ontology, Content & Attestation: knowledge in modular packages, organised by an optional ontology, its content kept with the original sources, and attestations, signatures that prove who published it and who checked it. Open Readers in TypeScript, .NET and Python read, verify, search and cite a package the same way, on a server or offline, and every answer can cite a source you can check.

Use MOCA with your AI assistantAdd the openmoca agent skill to Claude, Copilot or Codex, and ask it to make, sign, search and cite packages.

Got a .moca file?Open it in Claude Desktop, Claude Code, VS Code or Cursor with one download, and get answers that cite their source.

RAG with citations you can checkWhat a package adds to retrieval-augmented generation: current, in-scope content, every passage traced to its source.

Trust and provenanceSigned packages, signed reviews, quotes checked against the originals, and proof of what an answer used.

Grounding for AI agentsOne package for any agent: an MCP server, Agent Framework, LangChain and LlamaIndex, with the same trust rules.

Ontology for AI retrievalTopics, steps and prerequisites as structure, and search that adds what each result requires.

Sovereign and offline AIGrounding that runs on infrastructure you control: no network calls, local search, signatures you verify.

Questions?What MOCA stands for, how it relates to RAG, vector stores, OKF and MCP, and where to start.

The problem

Knowledge that feeds AI is often rebuilt for each model, vendor and pipeline, and loses its structure and its sources on the way. It is then hard to prove which version an answer came from, who published it, or whether anyone checked it. Superseded content can keep being served, because the pipeline is rarely told it is no longer in force.

Where it helps

MOCA is not another model, vector store or agent framework. It is the form your knowledge travels in, so whatever you build answers from the right version and cites the original source.

Sharing knowledge for groundingPackage a body of knowledge once; any assistant, model or framework grounds on it.

Manuals and proceduresSteps in order, the checks that come first, cited to the PDF page, even offline.

Law and regulationOnly the version in force, with the exact clause cited.

Courses, talks and videoLessons in order and prerequisites kept; every answer opens the video at the moment.

Customer support knowledgeAI drafts kept apart from what a person checked; internal answers stay internal.

Audit and regulated AIA record of exactly which content each answer used, long after the index is rebuilt.

More use cases, and where MOCA is a poor fit.

A package, in three layers

A MOCA package wraps an Open Knowledge Format bundle, a folder of Markdown with YAML frontmatter. The name maps onto it: the ontology is the structure layer, the content comes with its evidence, and attestation is the trust that runs through all three, like the pin in the logo.

Trust through all three

One digest covers every file of the package, so any change shows. Signatures and reviews sit beside it: a signed package proves who published it, and signed reviews prove who checked which node.

Structure optional

Concepts and how they relate: hierarchy, parts, order, what requires what. W3C SKOS and DCMI terms, plus any vocabulary you add, such as an OWL ontology, carried as it is.

Content

The knowledge, one Markdown node per topic, with validity windows, audiences and translations.

Evidence

The original sources, carried in the package where they can be, and selectors that point into them, so a quote can be checked against the original.

What a Reader gives you

A Reader opens a package, verifies it, and searches it with trust rules applied: content that is out of force, superseded or outside the caller’s audience never comes back. Every result is a citation record: the text, the package version and digest it came from, a versioned reference to cite, who signed and reviewed it, and its evidence, marked matched when the quote was found in the original.

Readers are in-process libraries, one per language, and every one passes the same conformance corpus, so a package reads the same way everywhere.

Get started in TypeScriptnpm install @openmoca/moca-core: install, open, search and cite, run the corpus.

Get started in .NETdotnet add package OpenMoca.Reader, with bindings for Microsoft.Extensions.AI, the Agent Framework and VectorData.

Get started in Pythonpip install moca-reader: the same Reader, for the language most AI code is written in.

Serve packages over MCPOne download, nothing to install: search, navigate and cite packages from any MCP client.

Make a packageA valid package in five minutes, from Markdown you already have.

Try it in your browser

Try OpenMOCA runs the .NET Reader over eight real packages, each converted from a published source and signed. No install, and no account to search.

• Learn Copilot Studio 2026 — a video series: 17 videos in order, every answer cited to the moment. MadeForCloud on YouTube, CC BY 3.0.
• First Steps to Using Claude AI — a recorded talk, its chapters cited to the moment. Eliot Prince on YouTube, CC BY 3.0.
• Kubernetes documentation — vendor documentation at scale: all 1,653 pages. The Kubernetes Authors, CC BY 4.0.
• Airplane Flying Handbook — a manual with a topic ontology: what each topic builds on. FAA, public domain.
• Data Protection Act 2018 — legislation, section by section. legislation.gov.uk, Open Government Licence.
• HSE health and safety guidance — regulated procedures, cited to the PDF page. Health and Safety Executive, Open Government Licence.
• GitLab handbook — onboarding: a company’s policies and people practices. GitLab, Creative Commons with attribution.
• Acme Support Knowledge Base — trust: 34 pages in English and French; internal procedures kept out of public answers, expired and future policies left out, stale and disputed pages flagged. The OpenMOCA specification’s example, Apache-2.0.

What you can do there:

• Search, with no account. Lexical (keyword), semantic or hybrid, through the Reader’s one search call. Two local embedding models, bge-small and Nomic, each with its own sidecar index of the same passages; your search is embedded on the server and never sent to an AI provider. Hybrid results show where each one ranked in the lexical and semantic lists.
• Open every result at its source. The passage opens beside the results, marked in place; a video opens at the moment.
• Use the structure. Search within a chapter, a folder or a topic, narrow to what the results fall under, or widen back out. In the Flying Handbook, each topic shows what it builds on (requires).
• Check, don’t trust. Inspect a package: its digest, its signer, the Reader’s findings and its search indexes.
• Ask it questions. Signed in, with your own key (Anthropic, OpenAI or Google Gemini) or a demo key by invitation. Every answer cites the package, and a citation opens the source.

The site is built and run by PI & Mash on the open Readers and tools; its own code is not part of OpenMOCA.

Three pillars

Pillar	What it is	Openness
MOCA Package	The format: structure, content, evidence, and the identity and trust that cover them	Open standard
MOCA Reader	The standard way to get anything out of a package: search, structure navigation, citation. In-process, one per language, one conformance corpus	Open source
MOCA Application	The product for one use case, built on a Reader and the agent framework of your choice	Yours

Status

OpenMOCA 0.6.0 is the first public release. The specification and three Readers, in TypeScript, .NET and Python, are released together and pass one conformance corpus. Until 1.0.0 the format may still change, and every change is noted; 1.0.0 comes when it has settled and is used beyond its maintainer. If you’d like to help get there, get in touch, or see the roadmap.