Prototype an MCP server exposing the pack as resources + a Blender-Python lint tool

[TMHSDigital]

Context / Problem Statement

CLAUDE.md and AGENTS.md both note "there is no MCP server." Today the content is consumable only by clients that natively read skills/+rules/ (Cursor, Claude Code). An MCP server would make the pack first-class in any MCP-aware client and turn the 6 .mdc anti-pattern rules into an executable lint tool — a far stronger guarantee than prose rules.

Proposed Solution / Implementation Steps

1. Add an optional adjacent package (e.g. mcp-server/) so the content repo stays a pure pack; the server reads skills/, rules/, snippets/ from the repo at runtime.
2. Expose MCP resources: list_skills, get_skill(name), list_snippets, get_snippet(name), search_patterns(query).
3. Expose an MCP tool lint_blender_python(source) implementing the rules as AST checks: bpy.ops.* inside loops, bmesh.new() without paired free(), prop assignment vs annotation, bpy.context.copy() override, per-element bulk loops. Reuse this checker in CI (a rules-as-linter long-term item).
4. Document install/usage in a mcp-server/README.md; keep it out of the count/validate gates (it's tooling, not pack content) but add its own minimal test job.

Definition of Done

• MCP server lists and serves skills/rules/snippets as resources.
• lint_blender_python flags all 6 anti-patterns on a fixture file and passes clean code.
• Server runs against the live repo checkout with documented setup.
• CI runs the server's own tests without touching the content-count gates.

---

Filed from the repository technical audit — see docs/technical-audit.md.