Flow MCP: smarter flow generation, validation warnings, and clearer preview limits

What changed

On 2026-08-07 we shipped a set of improvements to the Final Flows MCP server, the Flow project, and the Cursor plugin — all live now on flow.finalpos.com. The changes fall into four areas.

1. TypeScript types ship with the scaffold baseline

Before: The first time you ran validate_flow or synced a new scaffold, you'd hit a wall of TS7016 / TS7026 errors because @types/react and @types/react-dom were missing from the preview scaffold's devDependencies. The tool looked broken before you'd written a single line.

After: @types/react, @types/react-dom, and typescript are now part of the scaffold baseline and kept in sync with the seed template automatically. First-time syncs are clean.

2. validate_flow now returns advisory warnings

validate_flow has been upgraded to v2. It now returns a warnings[] array that surfaces four existing guardrail checks — previously only visible inside the in-app fix loop — directly to your MCP client:

• Client money math — flags flows that perform payment calculations client-side instead of on the server.

• Hardcoded catalog data — catches product IDs, prices, or SKUs baked into the flow code rather than fetched from the catalog.

• Missing `StationHomeButton` — warns when the station-home navigation anchor is absent.

• Missing `data-test` attributes — flags components that can't be targeted by automated tests.

• Tip/gratuity as cart fee — a new lint that catches the common mistake of adding a tip as a line-item cart fee instead of using the correct gratuity field.

validate_flow also now:

• Reports typechecker: ok or typechecker: unavailable explicitly, instead of returning a silently clean result when the type checker couldn't run.

• Stamps pinnedCommandFrameVersion and currentVersion in the response so you can see immediately if your scaffold has drifted from the deployed command frame.

3. deploy_flow blocks on hardcoded credentials

If your flow source contains hardcoded secrets or credentials, deploy_flow now refuses to proceed and returns an actionable error. Because a built flow can be distributed to other organisations, a leaked credential is an irreversible harm — this check runs before any deployment attempt.

4. start_preview emits a preview-limitations manifest

The preview environment has always had constraints that weren't documented anywhere an external agent could read them. start_preview now emits a version-stamped manifest that lists exactly what the preview does and does not simulate:

• Instant always-success tenders (no real payment processing)

• Canned card totals

• Cart appends that do not read back

• Query filters that are ignored

• The ?finoodle_mock dataset selector for switching mock datasets

This means your AI agent knows the preview's limits before it starts testing — not after it's confused by unexpected behaviour.

5. Rate budget disclosure in server instructions

The MCP server instructions now include per-key hourly rate budgets, interpolated directly from the enforcement constants. The prose and the enforcement can no longer drift apart.

Who benefits

These changes apply to every MCP client: Claude Code, Cursor, Gemini CLI, ChatGPT, and any other client connecting to the Final Flows MCP server. No configuration changes are needed on your end — the improvements are live in the server and scaffold baseline.

If you previously worked around the old behaviour with a custom skill or prompt file, you can simplify or remove those workarounds. The guardrails that required manual instruction are now built into the server.