The Pipeline

cl-mcp is a straight, one-directional pipeline. Understanding the four stages makes the whole system easy to reason about.

analyzer CLI → component-metadata.json (+ workspace-manifest.json) → server / CLI → LLM tools

Stage 1 — Analysis (build time)

The analyzer reads TypeScript/TSX source with the TypeScript compiler API and produces framework-agnostic metadata.

• Angular (analyzers/angular/) — components, directives, pipes, services, inputs, outputs, inheritance, content projection, deprecation and config tokens, driven by a ts.Program for full type resolution.
• React (analyzers/react/) — function/arrow/class components (including memo/forwardRef/React.FC<P> and custom factories), props via the TypeChecker (required/defaults/literal unions/JSDoc), /^on[A-Z]/ callbacks as outputs, children/ReactNode props as slots, and compound naming (Dialog.Root) from namespace barrels.

Both implement the same FrameworkAnalyzer interface and emit the same shapes.

Stage 2 — Metadata artifacts

Each library becomes one component-metadata.json. In multi-library mode a workspace-manifest.json records every library plus the cross-library import graph. These files are the contract between build time and serve time.

Stage 3 — Loading (serve time)

@cl-mcp/core’s data layer loads every resolvable metadata file into a registry keyed by library name, behind a Zod trust boundary. Domain code reads one “active” library at a time; handlers switch it per request. Paths resolve from CL_MCP_METADATA_PATH (single) or CL_MCP_DATA_DIR (all subdirectories) — a single file means single-library mode, fully backward compatible.

Stage 4 — Tools

The six tools are pure args → ToolResponse functions in core (TOOL_HANDLERS). The MCP server and the CLI are thin adapters that route to the exact same handlers — no domain logic is duplicated between them.

Design principle: no hardcoded library values

Nothing library-specific (selector prefix, package name, framework) is hardcoded in core or the server. It is all derived from the loaded metadata at runtime, so adding a library never means touching the serving code.