MCP server writes global KiCad library tables at startup and exits if it can't

[barbecuesteve]

Thanks for building this — the LCSC-to-KiCad conversion is just what I need.

One thing I noticed while evaluating it, in packages/mcp/src/index.ts:

async function main() {
  const registration = await ensureGlobalLibraryTables();
  if (!registration.success) {
    process.exit(1);
  }
  const httpPort = startHttpServer();
  ...
}

The server writes KiCad's global sym-lib-table and fp-lib-table before any tool is called. MCP hosts start servers speculatively — on session start, in every project, whether or not the user touches them. So merely having jlc-mcp in an MCP config mutates machine-global KiCad configuration that affects every project on the box, not just the one being worked on. For users with a deliberate library layout (shared repo-relative libraries, ${KIPRJMOD} paths) that's a surprising thing to have happen unasked.

The sharper problem is the process.exit(1). If the tables aren't writable, or the KiCad version isn't detected, the server doesn't start at all — so component_search, which is read-only and needs no libraries, becomes unavailable because of a filesystem write it never uses. A read path is gated behind an unrelated write.

Smaller version of the same thing: startHttpServer() binds a port unconditionally for the browser UI, so a user who only wants stdio tool calls gets a listening socket they didn't ask for.

Suggested fixes, roughly in order of value

1. Make registration lazy — call ensureGlobalLibraryTables() on first library_* invocation rather than in main(). No new config surface, and search-only users never touch the tables.
2. Degrade instead of exiting — on failure, log and disable the library_* tools, but keep serving component_search.
3. Add a --read-only mode that omits the five writing tools from the advertised tool list entirely. That's a stronger guarantee than skipping startup work, since a model can't call a tool that isn't listed — useful for anyone wanting sourcing lookups without granting write access to their KiCad libraries.
4. Same treatment for the HTTP server — --no-http, or start it lazily.

(1) and (2) alone would resolve it for me; (3) is the nice-to-have.

I'm happy to offer a PR for any subset of these if you'd welcome it. Just say which shape you'd prefer before I write anything, since the work spans packages/core and packages/mcp. Per your project protocol, I would include a changeset.

Unrelated: LICENSE file issue

package.json declares MIT, but there's no LICENSE in the repo, so GitHub reports the license as unset. Would you consider adding one? It makes contributing unambiguous.