Code Knowledge Bundles¶
Turn a source repository into knowledge an agent can act on — where every claim points back at the code it came from.
A Code Knowledge Bundle is an OKF bundle built from a
repository rather than from documents. codekb parses the repository, resolves the
relationships between its symbols, extracts claims about what the code does, and
records the exact source span behind each one.
The result is a folder you can commit, diff, and hand to an agent:
- Every claim traces back to the lines it was derived from.
- Nothing reaches an agent until the tier that needs review has been reviewed.
- Built on OKF's core model, so it is not a proprietary export format.
What To Read First¶
| If you want to | Read |
|---|---|
| Build your first bundle | Getting Started |
| Choose between static and hybrid mining | Pipeline & Mining |
| Know what a claim is and when it counts | Knowledge & Review |
| See what is detected for your stack | Architecture Detection |
| Wire a bundle into an agent | Agent Consumption |
| Expose a bundle over MCP | MCP Server |
| Know how the package is put together | Architecture |
| Look up a command | CLI Reference |
How A Bundle Is Built¶
flowchart LR
R[Repository] --> S[Scan]
S --> P[Parse]
P --> RE[Resolve relations]
RE --> A[Detect architecture]
A --> M[Mine claims]
M --> RN[Render human layer]
RN --> V[Validate]
V --> B[(Code Knowledge Bundle)]Each stage writes its own state file under .codekb/, so a run that fails partway
leaves everything the earlier stages produced. See
Bundle State.
Where This Sits¶
KL4A is the product; codekb and sopkb are the two options it offers.
codekb is the code option: a repository in, grounded claims out. sopkb is the
SOP option — Word docs, PDFs and Markdown in — and has its own tab.
Both write OKF bundles over one shared core, gate retrieval behind human review, and expose the result over a CLI and an MCP server. They differ in what goes in and what gets extracted.
They are separate installable packages. See Architecture.
Languages¶
codekb parses Python and COBOL. Files in other languages are inventoried
and counted, so you can see what a repository contains, but they are not parsed
into modules, symbols and relations — nothing is claimed about them.
What This Is Not¶
- Not a code search tool. Search exists, but the point is the reviewed claim layer on top of the parse, not the parse itself.
- Not a linter or a static analyser. Nothing here judges your code. It records what the code contains and what can be said about it with evidence.
- Not a substitute for reading the code. A bundle is a map. Every entry links to the territory.