Install

ContextCake is clone-and-run. The engine has zero dependencies — there is no npm install step, so there are no install scripts, no transitive packages, and no lockfile to trust. What you clone is what runs, and you can audit all of it.

Prerequisites

• git
• Node.js ≥ 18

1. Clone

2. Resolve your first concept

The repo ships a tiny three-layer demo bundle. Resolve a concept where the layers disagree and watch the merge — provenance and conflicts included:

cd context-cake
node resolver.mjs --manifest playground/manifest.json --concept decisions/primary-db

You'll get JSON: each section tagged with the layer that won it, and a conflicts array where the losing layers dissent — with dates.

3. Serve it to an agent

Start the MCP server against the same demo bundle:

node mcp-server.mjs --manifest playground/manifest.json

Or register it with an MCP client (Claude Code shown):

claude mcp add contextcake -- node /path/to/context-cake/mcp-server.mjs --manifest /path/to/context-cake/playground/manifest.json

From here, point the manifest at your own layers — see Your first cascade and Connect an agent.

Optional: explore visually

npm run playground   # http://127.0.0.1:8790

The playground is a local canvas over the same engine — sources, live merges, and an N-way conflict resolver. (It runs from the clone too: its libraries are vendored in the repo, not fetched.)

A note on trust

The manifest is a trust boundary: a layer with "source": "mcp" spawns the command it names, as your user. Only point ContextCake at manifests you wrote or trust — the same rule as any MCP client config. Read the details.

Why isn't this on npm?

Deliberately, for now. Registry supply-chain attacks have repeatedly hit AI and agent tooling — compromised maintainer accounts shipping credential-stealing postinstall payloads to thousands of downstreams. A knowledge engine that 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.