# llmdoc

Persistent engineering context for coding agents, so an agent retrieves only the decisions, constraints and cross-module contracts a task needs, verifies them against the live repository, and rechecks them as code evolves instead of rereading everything each session.

llmdoc stores durable engineering meaning, not a copy of the repository, under a `llmdoc/` directory; an agent first retrieves the smallest useful knowledge set for a task, uses the source and tests to confirm current facts, and later verifies affected knowledge as a review obligation, without automatically rewriting documentation on every code change.

For Claude Code, it installs as a plugin through `/plugin marketplace add` and `/plugin install`, then initializes a repository with an `/llmdoc:init` skill; for Codex, it adds through `codex plugin marketplace add` and the `/plugins` panel, then runs the same init skill from a prompt. A repository that already has an `llmdoc/` directory can be queried directly with `npx -y @tokenroll/llmdoc tree` or `search`, without installing the plugin.

The external `@tokenroll/llmdoc` package should not be added to a consuming project's own `package.json` or lockfile, and pinning a version with `npx -y @tokenroll/llmdoc@` is recommended for reproducible runs. It is written in TypeScript and needs Node.js 18 or newer and a Git repository.

Website: https://github.com/TokenRollAI/llmdoc