Add CanMCP endpoint diagnostics skill

[OlegNickeshin]

Summary

Add CanMCP, an agent skill for diagnosing public remote MCP endpoints with the existing local Python CLI.

Typical intent: Diagnose why this MCP endpoint fails to connect, without executing any server tools.

• Source: https://github.com/OlegNickeshin/canmcp
• Public skill: https://github.com/OlegNickeshin/canmcp/blob/main/skills/canmcp/SKILL.md
• Raw skill: https://raw.githubusercontent.com/OlegNickeshin/canmcp/main/skills/canmcp/SKILL.md
• Existing CLI distribution: https://pypi.org/project/canmcp/

Scope

This entry is application/ai-skill, not an MCP server. CanMCP remains a CLI requiring command execution, Python 3.11+ and direct DNS/HTTP(S) access. There is no hosted scanner or remote connector endpoint.

The skill explains how to run a bounded scan, read JSON and nonzero exit codes, distinguish protocol failures from incomplete inspections, and preserve the scanner's public-target/TLS boundary. Normal inspection does not sign in, register clients or call server tools. Optional OAuth requires user authorization; Dynamic Client Registration may leave a provider-side registration. The browser for the loopback callback must run on the CLI machine. Results are local diagnostics, not conformance certification or proof that a particular ChatGPT/Claude account can connect.

No endpoint credentials, private targets, conversation content or tokens are published. Secret path segments require redaction before sharing reports.

Validation

• Skill frontmatter/scaffold validation passed.
• CLI flags and installed version checked against CanMCP 0.2.0.
• Full CanMCP suite: 201 tests passed, including scripted protocol and local TLS/OAuth fixtures.
• Ruff lint/format checks passed; sdist and wheel build succeeded. The skill is included in the source archive.
• Default CLI inspection of the owner's public PeopleMCP endpoint completed without OAuth or tool execution. Its diagnostic profiles passed; this is not a cloud-client acceptance test.
• Agent Finder tests: 16 passed.
• python3 scripts/generate_ai_catalog.py and --check passed.

Only catalog/OlegNickeshin/canmcp.json is committed here. metadata.sourceSet is OlegNickeshin/canmcp, as required by the validator. Regenerating ai-catalog.json also changes unrelated live upstream metadata (1,844 diff lines); it was validated locally but omitted from this focused source-entry PR. The documented main-branch fallback can regenerate the ingestion artifact after merge. No generator or workflow changes.