Atoms distills durable knowledge from your development process — Claude Code
sessions, GitHub PRs, and commits — into a reviewed atom store (.knowledge/),
then projects it into human docs and AI-context files. You review; you don't author.
Documentation is a derived artifact. A machine distills the knowledge that already
lives in your sessions, PRs, and commits, and projects it into whatever form a
reader — human or agent — needs.
Features
Distill sessions, commits, and PRs into atoms — one durable fact each, with a
claim, subjects, and provenance back to the source.
Reconcile new evidence against the store: merge, supersede, relate, or create —
so the same fact never spawns duplicates.
Project atoms into docs and .claude/ files inside atoms:begin/atoms:end
regions, so generated content and your hand-written prose coexist without clobbering.
Review-gated: nothing is written silently. Accept, edit, or reject each proposal.
Auto-triggers: on session end, on each commit, and (optionally) on merged PRs.
Zero-config LLM: defaults to the claude CLI you already use; or bring an
Anthropic API key.
Privacy first: secrets are redacted before anything reaches the model.
Getting started
Install the extension and open a project that has Claude Code sessions.
Open the Atoms view from the activity bar.
Run “Atoms: Review & Distill a Session…”, review the proposed atoms, and apply.
Your atoms land in .knowledge/ (commit it to share with your team) and render into
the projection docs.
Settings
Configure the store path, LLM backend, trigger cadence, and redaction patterns under
the Atoms section in Settings. See the
configuration reference.
Also available as a CLI
The same engine ships as a CLI: npm install -g @bebridge/atoms for use in a terminal,
CI, a git hook, or a Claude Code Stop hook.
