Quickstart

Lay out the source

Point langonrock at a folder where every immediate subdirectory is one bundle.

sources/acme/
  sales/
    tables/orders.md
    tables/customers.md
  ops/
    runbooks/deploy.md

sales and ops are bundles. orders, customers and deploy are concepts. Frontmatter is optional: a plain Markdown file still compiles, with its id from the path, its summary from the first sentence, and its title from the first heading — it just earns a conformance warning, which --strict turns into a failing exit. Only navigation files, index.md and log.md, never become concepts.

Compile it into the store

langonrock sync sources/acme --data ./data --tenant acme

snapshot eaf251169a59 (new), 2 bundles [ops sales], 3 concepts, 850 bytes on disk

That is the whole write path. sync reads every subdirectory as a bundle and writes one snapshot for the tenant.

Read the manifest

langonrock manifest --data ./data --tenant acme

# tenant: acme
# bundles: ops sales
id	bundle	kind	status	grain	summary	links
deploy	ops	runbook	-	-	How to ship the orders service.	-
customers	sales	bigquery_table	deprecated	customer_id	Registered customers, including churned.	-
orders	sales	bigquery_table	-	order_id	One row per completed customer order.	customers

Everything an agent needs to plan its reads is in those rows. The links column carries the graph, so it knows every id it will need before it fetches anything. A status cell other than - is the concept telling you it is not current — deprecated, draft, or stale once its stale_after date passes, which the reader computes at read time so the snapshot itself never depends on the clock.

Fetch only what you chose

langonrock get orders --section schema --data ./data --tenant acme

Section names come from the concept's own Markdown headings, lowercased with non-alphanumerics collapsed to underscores. A # Common query patterns heading becomes --section common_query_patterns. Text before the first heading is the body section.

Drop --section for the whole concept, and pass several ids to fetch them in one call. When you know a phrase rather than a section, --find "exact words" returns the window around it plus the offset of every occurrence — the cheap way into a long document with no headings.

Keep it following your edits

Syncing by hand gets old. Run the watcher instead:

langonrock watch sources/acme --data ./data --tenant acme

Add a bundle by creating a folder, remove one by deleting the folder, change one by saving a file. The watcher debounces filesystem events at 200 ms and runs a full rescan every 30 seconds as a backstop, because watch events are a hint on every platform and never the source of truth.

If you are going to run the daemon anyway, skip this command. serve starts the same watcher itself, and running both would put two watchers on one tenant.

Hand it to an agent

langonrock mcp "okf://$PWD/data?tenant=acme"

For Claude Code, register it once:

claude mcp add langonrock -- langonrock mcp "okf:///abs/path/to/data?tenant=acme"

Six tools, no more: manifest, search, get and snapshot to read, write and delete to change. See MCP server for what each one is for and why the count is capped.

Searching instead of reading everything

langonrock query "okf:///abs/path/to/data?tenant=acme" search orders

Search returns manifest rows, never bodies, so the two-hop path holds. Narrow with search, then fetch the ids you chose with get. Results include concepts one link away from the top matches, which is usually where the join partner or the parent dataset lives. Each direct hit also ends in a pos column — the offset where your query's words cluster densest in the body — so the second hop can be get with --offset <pos> --limit 2000 instead of the whole document, even when there is no literal phrase for --find to anchor on.

Progress goes to stderr

Stdout stays clean for piping, and Bun paints anything written with console.error red. Those lines are status, not failures.