copilot-money-mcp: Instructions file for Claude Code

Copilot Money MCP Server

MCP (Model Context Protocol) server that enables AI-powered queries and management of Copilot Money personal finance data. Reads come from the locally cached Firestore database (LevelDB + Protocol Buffers); writes go through Copilot's GraphQL API at app.copilot.money/api/graphql. 33 base tools (14 read + 19 write); --live-reads swaps 6 cache reads for 17 live tools (25 read tools in live mode). Counting convention: "base" = cache-mode read tools (createToolSchemas()) + write tools (createWriteToolSchemas()); live-mode _live variants are counted separately. Read-only by default, write tools opt-in via --write flag.

Quick Reference

bun install          # Install dependencies
bun test             # Run tests
bun run build        # Build for production
bun run pack:mcpb    # Create .mcpb bundle for Claude Desktop
bun run check        # typecheck + lint + format:check + check:version-sync + check:server-json + check:deps-pinned + check:docs (check:tool-counts, check:bug-corpus, check:privacy-endpoints) + check:concealment + check:workflows + check:tracked-files + check:action-pins + check:mutation-guards + bun test --bail
bun run fix          # Run lint:fix + format

bun run check does NOT run check:ghost-lines either. That mode of scripts/check-concealment.ts compares a PR's commits against its final tree, so it needs a commit range rather than a working tree; it runs as its own pull_request job in test.yml (with fetch-depth: 0) and can be run by hand with bun run check:ghost-lines. With no range it prints SKIPPED and exits 0 locally, and exits 1 in CI.

bun run check does NOT run check:skills (the skills/ linter). Run bun run check:skills separately when touching anything under skills/. It shells out to scripts/dump-tool-names.ts and scripts/dump-tool-args.ts under bun to get the real tool list and each tool's argument names, so it needs bun on PATH and a completed bun install. It also cross-checks the field names skills reference against each terse-by-default tool's DEFAULT_*_FIELDS preset (#704), so a skill that reads a field the v3 diet dropped fails the lint instead of failing silently at use time. These checks, and a table-shape check (every row has as many cells as its header), read every .md file under a skill's folder, not just SKILL.md — so a skill's test matrix or reference files are linted too. Every skill declares metadata: modes: "live, cache" (or "live") in its frontmatter, and the lint fails a tool or argument those modes don't list (#847); a two-mode skill names both variants, tagged: `get_accounts_live` (live) / `get_accounts` (cache).