SpandrelA framework that turns markdown file trees into governed knowledge graphs — philosophy, spec, and reference implementation

Deployment

Four deployment modes — local in-memory dev, static + flat-file MCP for read-only publishing, hosted live backend for writes and identity-aware reads, and Claude Code plugin packaging for delivering a graph + skills + MCP as a single installable unit

Deployment

Spandrel runs in four modes with different infrastructure needs. The same wire surfaces (MCP and REST) are served in all of them — the only difference is where the compiled graph lives and who reads from it:

  • Local development — single-process dev server with the in-memory store; zero setup beyond spandrel dev
  • Static + flat-file MCP — spandrel publish emits a static bundle; a thin serverless function translates MCP over the bundle. Read-only, hostable anywhere, embeddable in an existing site
  • Hosted live backend — compiler writes to a persistent store; REST + MCP serverless; graph updates on push. Appropriate when the graph needs writes, identity-aware reads, or federation across repos
  • Claude Code plugin — package the graph + MCP server + skills as a Claude Code plugin installable via the marketplace mechanism. Best for delivering client-specific knowledge as a sealed, versioned unit consumed by an agent inside Claude Code