A local-first system of record and effectiveness monitor for your agentic
assets: skills, subagents, MCP servers, and bundles. Point it at your own
~/.claude, and every asset you have built carries evidence of use, cost,
measured effectiveness, and governance, so the best assets improve and the
stale ones retire, with your usage data never leaving your machine. The full
privacy contract lives in docs/PRIVACY.md.
Built with React 19, Vite, and Tailwind on a Microsoft Fabric-native data model (TypeScript entity decorators). Runs entirely on deterministic seed data out of the box; wire your own data with the bundled collector.
git clone https://github.com/cdrguru/living-vault.git
cd living-vault
npm ci
npm run dev # local UI on synthetic seed dataOpen the printed localhost URL. Everything you see is generated demo data.
The standalone collector scans your machine locally and writes JSON you can upload on the Settings page (browser-only; nothing leaves your machine):
node scripts/living-vault-collect.mjs assets # your skills/agents/commands -> projection JSON
node scripts/living-vault-collect.mjs telemetry # your Claude Code usage -> upload JSONCollectors record metadata only, never message or command content. See docs/PRIVACY.md for the full privacy contract.
| Surface | What it shows |
|---|---|
| Monitor | Operational health, exception queue, per-harness usage |
| Effectiveness | Ranked assets with evidence-graded confidence |
| Cost | Spend by harness and model, wasted-spend estimate |
| Coverage | Coverage matrix and duplicate clusters |
| /model | Interactive entity-relationship canvas of the data model |
The architecture story lives in docs/site/ARCHITECTURE.md.
npm run lint
npm test # public-scope Vitest suite
npm run build:swa # full static bundle: landing at /, demo at /demoContributions are welcome. Start with CONTRIBUTING.md;
the short version is: fork, branch, npm test green, open a PR.
This repository is a curated public export of a private upstream. Generated
files (most of src/) are refreshed from upstream and PRs against them are
cherry-picked back rather than merged directly; community-owned files (this
README, docs, templates) are yours. The export tooling marks which is which
in .public-export/state.json.
MIT. See LICENSE.