The Agent
The Agent is the part that reads. It reads the documents, interviews you about the vocabulary, writes one proposal per document with the sentence for each fact, answers questions with the document and the date, captures corrections said in conversation, and writes specs whose every claim cites the graph. It is a set of skills and tools for Claude Code, where OTO is tested, or for the harness you prefer, since the skills are plain files and the tools a standard MCP server; the agent brings its own model, and the engine gates everything it proposes.
pluginthe engine plugin, installed from the registry or with any pack10 skillsbuild-knowledge-base, ontology-interview, curate, capture, query-knowledge, spec, evaluate, vet-provenance, act, concierge14 toolsthe kg_* query tools as an MCP server for the open projectpacksa domain's ontology, skills and views, with a start skill; four in the catalogoto drafta model drafts one document's proposal, to the same rulesdomain skillsskills that ship in packs: the report-automation pack brings fivethe factoryan agent that builds, tests and runs from the graph as its specificationWhat it does
| When | Skill | What the agent does |
|---|---|---|
| a folder of documents | build-knowledge-base | ingests, surveys, designs the vocabulary, proposes each document, stops at three gates, serves the result |
| the vocabulary | ontology-interview | asks what the graph must answer, proposes classes with a reason each, records who confirmed |
| one document, or a change | curate | writes the proposal, merges through the dry run, shows the diff, applies by name |
| something said | capture | writes the statement as a dated source in the speaker's words |
| a question | query-knowledge | resolves the terms, reads current facts, cites the source and the date, says what the graph does not know |
| a spec or a decision | spec | writes it from the graph, every claim cited, decisions in a shape that becomes facts |
| what can be done | act | shows an action bound to an entity for the caller; records the response |
| where am I | concierge | quotes the project's state, explains the station, names the next playbook |
The Verifier's two skills, evaluate and vet-provenance, are the agent checking its own work; they are listed under the Verifier.
The rule it works under
The agent proposes; it never applies. Every proposal passes the dry run, the merge, the check and apply by name, the same gates a person's proposal passes. A model's mistake is caught the same way as a typo. The engine never calls a model to build, gate or answer.
Its pages
Install in Claude Code. The skills, one page each from the skill's own playbook. Packs. The catalog.