Onboard an AI agent to a knowledge base with okf prime

Onboard an AI agent to a knowledge base

An agent starts with the map of your catalogue and the pages that matter, in two calls.

Search terms covered: onboard AI agent to knowledge base, okf prime, Claude Code context, MCP knowledge base. Output below comes from okf 0.2.5 against a copy of the golden test catalogue; <catalogue> replaces the real path.

Set it up

Install okf, then connect your agents. See Installation for the install paths.

okf install            # preview: prints a plan, changes nothing
okf install --apply    # installs skill, MCP server entry, completion

Restart the agent afterwards. The MCP server runs the okf binary; OKF_BIN overrides which one. See Agent onboarding.

Work through it

1. Give the agent the map. okf prime prints the usage guide, the command directory and the catalogue state with the bundle TOC. An agent runs it once at session start.

## This catalogue

root: <catalogue>
bundles: 4 · concepts: 24
fts index: absent — `find` degrades to a substring scan; run `build-index`
semantic index: absent — `find` runs its lexical arm only
tag vocabulary: closed — 11 tags in TAGS.md; check a new one with `tags --validate`, adopt it with `tags --register`
ctx bundles: 3 — cooking · legacy · onboarding — `ctx show <slug>` lists one
root docs: SPEC.md · TYPES.md · PROPERTIES.md · TAGS.md

## Bundles
[exit 0]

2. Hand over one topic. okf ctx lists the context bundles. okf ctx <slug> prints that topic’s pages in reading order, with a token estimate.

ctx onboarding — Read this first on day one

required — read in this order:
  1. <catalogue>/handbook/getting-started.md  [Guide] Getting started  ~158 tok
     First steps: accounts, tools, and where to find things
  2. <catalogue>/handbook/glossary.md  [Concept] Glossary  ~62 tok
     Terms used here: "bundle" and "concept"
  3. <catalogue>/handbook/ops/deploy.md  [Runbook] Deploy  ~150 tok
     How to deploy: build, ship, verify
reference — query when needed:
  - <catalogue>/handbook/security.md  [Gadget] Security basics  ~55 tok
     why: Background on passwords

4 entries (3 required, 1 reference), ~370 tokens for the required set
[exit 0]

3. Let the agent look things up. Through MCP it calls the tools okf_find and okf_read. In a shell it runs okf find and okf read. A section read returns only that section:

## Steps

Mix, rise, bake. Related: [pasta](pasta.md).

What you get

An agent that finds a concept with ranked search and reads one section of it, with no blind grep over files. Curate more topics with Context bundles.

Limits

The prime state text, like tags --register, names build_tagvec.py, which a Go install lacks [erpf]. Ignore it, or run okf build-tagvec --produce. See Known limitations.

Onboard an AI agent to a knowledge base with okf prime