fluencydrift-mcp

MCP server that diffs a project's actual code structure against the architectural decisions captured in its fluencyloop store, and flags where reality has drifted from the record.

Status

Working end-to-end: reads a fluencyloop store, checks it against the live repo, and exposes list_architectural_records, check_drift, and explain_record over MCP. Built with fluencyloop itself — see docs/fluencyloop/ for the plan and feature history.

Related MCP server: docverity

Development

npm install
npm run build   # compiles src/ to dist/, and makes dist/index.js executable
npm test        # unit + in-process end-to-end tests (node --test)
npm run smoke   # dogfood smoke test: runs the built dist/index.js as a real
                 # subprocess against this repo's own fluencyloop store

Registering with Claude Code

After npm run build, register the built server as an MCP server:

claude mcp add fluencydrift-mcp -- node "$(pwd)/dist/index.js"

This registers it at the default local scope (private to you, in the project you run it from). Verify with claude mcp list; remove it with claude mcp remove fluencydrift-mcp. Point Claude Code at any fluencyloop-tracked repo's root as the working directory when you launch it — fluencydrift-mcp reads the store and scans reality relative to process.cwd().