mcp-per-projectlisted

# Per-project MCP Also run as **step 6 of `onboard-repo`** for every product backend or sibling app — do not wait for the user to say "MCP". ## Problem Both agents have a machine-wide MCP scope that every session loads. Putting `railway` + `railway-acme` there means Globex sessions see Acme tools, and a single `railway login` fights across accounts. | | Machine-wide (keep empty of product infra) | Per-project (use this) | |---|---|---| | **Cursor** | `~/.cursor/mcp.json` | `<repo>/.cursor/mcp.json` | | **Claude Code** | user scope, `~/.claude.json` | `<repo>/.mcp.json` (project scope) | Claude Code also has a *local* scope — project-specific but private to you, stored outside the repo. Use it for a server you need but shouldn't commit; `.mcp.json` is the checked-in, team-shared one. ## Rules 1. **Machine-wide scope stays free of product infra.** Cursor: `mcpServers: {}` in `~/.cursor/mcp.json`. Claude Code: add nothing with `--scope user`. Exception: a genuinely universal, account-agnostic tool. 2. **Each product defines servers in its own repo** — `.cursor/mcp.json` for Cursor, `.mcp.json` for Claude Code. Commit `.mcp.json` and `.cursor/mcp.json.example`; gitignore `.cursor/mcp.json`. 3. **Name servers with a product suffix**: `railway-acme`, `railway-globex`, `railway-postgres-acme`, … Never a bare `railway` when several products share the machine. Note Claude Code reserves `workspace`, `claude-in-chrome`, `computer-use`, `Claude Preview`, `Clau