MCP Pipeline Builder

Purpose

Automate the path from "I have a repo" to "I have a working MCP server registered with Claude Code." The pipeline derives what to expose by analyzing the repository — no prior knowledge required from the user.

Input: A repository URL or local path (e.g., https://github.com/org/repo or /home/user/my-tool). Output: A compiled, registered MCP server entry in ~/.claude.json or .claude/settings.json.

Operator Context

Note on context: fork: This pipeline intentionally does NOT declare context: fork, unlike other multi-phase pipeline skills. The Phase 2 human review gate requires interactive back-and-forth with the user; forked context would isolate that interaction from the parent session. The trade-off (larger parent context) is acceptable given the Phase 2 gate is the pipeline's most important safety mechanism.

Hardcoded Behaviors (Always Apply)

• Human Review Gate at Phase 2: ALWAYS pause after producing design.md and show it to the user before generating any code. The user must explicitly say "yes", "proceed", or "y" before Phase 3 begins. This gate exists because Phase 1 analysis is imperfect — catching a wrong tool name costs seconds; fixing generated code costs minutes.
• Read-Only Default: Only read operations are mapped to tools unless the user passes --allow-destructive. Phase 1 must classify every operation as read or write. Phase 2 silently drops all write operations from the design unless --allow-destructive is set.