| name | tooling-mcp-servers |
| description | Lazy MCP discovery via mcporter. Hard-trigger before any curl, gh, or HTTP fallback. |
mcp-servers
Triggers
mcp, mcporter.
The lazy-loaded MCP layer. Configured servers live in <personal-os-repo>/config/mcporter.json (symlinked from the workspace root so mcporter finds them via its default ./config/mcporter.json lookup). Typed headers per server live in <personal-os-repo>/mcp-servers/*.d.ts.
The point of this layout: discovery is cheap (this skill + the per-server index) and schema is paid only for the server actually needed. This is the lazy path - the only path on Codex, and the fallback anywhere. Note the harness split: on Claude Code the same merged inventory is also registered natively into user-scope (mcp__<name>__* tools) by sync-claude-mcp.py, since Claude's 1M window + on-demand schema deferral make native registration near-free. Both paths reach the same servers. See Design notes ("Eager vs lazy is now a per-harness split").
Hard-trigger rule
If the agent is about to reach for anything that smells like an MCP-shaped capability (search a corpus, drive a browser, query an observability backend, hit a service over HTTP that has an MCP wrapper, etc.), this skill fires first. Falling back to curl, gh, raw HTTP, or "I'll just shell out" without checking the mcporter inventory is the bug. The inventory (or mcporter list when the inventory drifts) enumerates everything flat - read it before improvising.
Sections
