Build an MCP server exposing KShield's scan engine as agent tools

[mesrikanthreddy]

Why

KShield's value today is gated behind the git pre-commit hook. AI coding agents (Claude Code, Cursor, etc.) can't call the scan engine directly mid-session — they only see results after a commit is blocked. An MCP server would expose the existing local backend as callable tools, letting an agent check code for secrets/access-control issues/hallucinated deps before writing more code, not just at commit time.

Proposed scope

• New package, e.g. mcp-server/ (Node, using the official MCP SDK), acting as a thin proxy over the existing FastAPI backend — no engine logic duplicated.
• Tools to expose:
  • kshield_scan(filename, content) → wraps POST /api/v1/scan
  • kshield_history(limit) → wraps GET /api/v1/scans
  • kshield_suppress(rule_type, justification) → wraps POST /api/v1/actions/suppress
• Needs a decision on backend lifecycle: should the MCP server auto-spawn kshield-backend if not already running, or fail fast with a clear "run kshield init" message? Leaning toward fail-fast for v1 to avoid managing a child process.
• Package for both npm (npx kshield-mcp) and inclusion in the existing CLI distribution channels (homebrew/pip) as a follow-up.

Out of scope for v1

• Auto-remediation via the agent (apply-patch endpoint) — revisit once read-only scanning is solid.