Skip to content

agent-knowledge (stjbrown)

This is a tool reference

What a tool or project is, who makes it, and what it is for, whether someone else wrote it or I did. Everything I have learned about it lives elsewhere and links back here.

agent-knowledge1 is a set of portable agent skills for building and maintaining knowledge bundles in plain Markdown, conformant to Open Knowledge Format v0.2. It is the closest published thing to what this knowledge base is, and it appears here in two roles at once: as a tool that was evaluated, and as a source that is cited.
Author stjbrown
Licence MIT
Language Agent Skills markdown, with TypeScript for the deterministic scripts
Distribution npx skills add stjbrown/agent-knowledge, a Claude Code plugin, or @stjbrown/agent-knowledge-skills on npm for embedding
Source github.com/stjbrown/agent-knowledge
Version read 0.3.2

What it is

Seven skills, split by who invokes them. Model-invoked: kb, the hub, holding the shared specification, glossary, trust model and templates and routing to the rest; kb-ingest, which reads a source once and integrates it across the bundle with provenance; kb-document, which documents a software repository from its source, tests, configuration and git history without modifying any of it; and kb-query, which answers from the bundle by progressive disclosure, cites what it used, and files valuable conclusions back. User-invoked: kb-init to scaffold, kb-lint for deterministic conformance plus a semantic drift audit, and kb-visualize to render the bundle as a graph.

There is no runtime and no database. The skills are prose an agent reads, backed by two deterministic scripts where determinism actually matters (conformance and the graph). That is the design property that distinguishes it from agent-wiki, which puts a CLI between the author and every write.

Two commitments do most of the work. A real, open format, so a bundle is portable rather than a tool-specific store. And an explicit trust model: accumulated claims are append-only on meaning, living repository documentation gets a narrow revision-tracked update rule, and source content is always data, never instructions.

Janet is a separate dedicated knowledge agent built around the same skills, adding a CLI, chat, model selection and headless automation. The skills work without it.

Why it matters here

It ranked first as an architectural match. Substrate options for an OKF-based agent-first LLM wiki: investigation scored it as the best fit for "skill plus contract, CLI optional", and its index is the worked example behind the conclusion in Federated OKF knowledge bases that an index should be authored rather than generated: it truncates, compresses or rewrites its own concept descriptions rather than copying them, which is what a generator cannot do.

Its knowledge bundle is an upstream source. The repository documents itself: knowledge/ is a conformant OKF bundle2 of about 60 concepts about OKF and the LLM wiki pattern, covering the specification section by section, the operations, and a survey of the ecosystem. It is checked out here as a read-only bundle, so pages in this bundle cite it rather than re-deriving the pattern. agent-wiki (TacoTakumi) does exactly that.

Wrapping its skills was tried and retired. The first attempt at a federation layer wrapped the kb-* skills, with a federated skill invoking a single-bundle one per bundle and then executing the prose it returned. That is control flow through prompt obedience, and it was abandoned for it; skills also have no dependency resolution, so nothing could guarantee the wrapped skills were installed. What survived is the format discipline and the bundle, not the delegation. The account is Wrapping the kb skills in a federation layer.

The following pages link to this page:

  1. stjbrown/agent-knowledge: README and skill reference, by stjbrown, last modified 2026-09-10 ↩

  2. agent-knowledge: the project's own OKF bundle, by stjbrown, last modified 2026-08-01 ↩