Installation
ContextCake is clone-and-run. The engine is dependency-free: there is no
npm install step, no install scripts execute, and nothing is fetched beyond the
clone itself. What you clone is what runs.
Prerequisites
Section titled “Prerequisites”git- Node.js ≥ 18
git clone https://github.com/siracusa5/context-cake.gitcd context-cakeThat’s the whole installation.
Verify
Section titled “Verify”Resolve a concept from the bundled three-layer demo, where the layers deliberately disagree:
node resolver.mjs --manifest playground/manifest.json --concept decisions/primary-dbThe JSON output shows the effective merge: contributors lists each layer with its
last-updated date, every section carries the sourceLayer that won it, and sections
where layers disagree carry a conflicts array with the dissenting layers’ content
and dates — surfaced, not hidden.
To run the full test suite (requires bash):
npm testWhat you just installed
Section titled “What you just installed”| Piece | Run with |
|---|---|
| Cascade resolver (CLI) | node resolver.mjs --manifest <manifest> --concept <id> |
| MCP server for agents | node mcp-server.mjs --manifest <manifest> |
| Interactive playground | npm run playground → http://127.0.0.1:8790 |
| Capture write path | node ingest.mjs / node write.mjs |
Why there’s no npm package
Section titled “Why there’s no npm package”Deliberate, for now. Package registries have repeatedly shipped compromised
AI/agent tooling through hijacked maintainer accounts and postinstall payloads.
A knowledge engine your agents read from should have a supply chain you can audit
in an afternoon — a git clone at a commit you can review is exactly that. If a
package is published later, it will be announced in the
changelog with provenance attestation.
- Your first cascade — build your own layers
- Connect an agent (MCP) — wire it into Claude
- The trust boundary — read this before pointing a manifest at sources you didn’t write