Flow MCP: first-time flow generation just got more reliable
The Final Flows MCP server now surfaces validation warnings to your AI tool, ships TypeScript types with the scaffold, blocks deployment on hardcoded credentials, and documents preview limitations, so you get reliable results from the first prompt.
Smarter flow generation, out of the box
We have shipped a set of improvements to the Final Flows MCP server that make AI-assisted flow creation more reliable from the very first prompt: no custom skills, no workaround prompt files, no wall of TypeScript errors on first sync.
Here is what is new.
TypeScript types in the scaffold baseline
The preview scaffold now ships with @types/react, @types/react-dom and typescript in devDependencies by default. First-time syncs are clean, and the TS7016 errors that greeted every new MCP user are gone.
validate_flow now talks to your agent
Final Flows already ran four guardrail checks on every flow: client-side money math, hardcoded catalog data, a missing Station Home button, and missing test attributes. Until now those checks only fed the in-app fix loop. External agents got a clean bill of health while shipping the exact problems the scans detect.
With this update, validate_flow returns a warnings[] array that surfaces all four checks, plus a new tip and gratuity lint, directly to your MCP client. Your agent sees the issues and can fix them before you ever deploy.
validate_flow also stamps the pinned and current command-frame versions in the response, and reports typechecker: ok or typechecker: unavailable explicitly. No more silently clean results.
deploy_flow blocks on hardcoded credentials
If your flow source contains a hardcoded secret, deploy_flow refuses to proceed. A built flow can be distributed to other organisations, so a leaked credential is the one mistake you cannot undo. The check runs before any deployment attempt.
Preview limitations, documented where your agent can read them
start_preview now emits a version-stamped manifest of what the preview environment does and does not simulate: instant always-success tenders, sample card totals, cart appends that do not read back, ignored query filters, and a selector for switching mock datasets. Your agent knows the preview's limits before it starts testing, not after it is confused by them.
Who this is for
Anyone building flows with their own AI tool over MCP: Claude Code, Cursor, Codex, ChatGPT or any Streamable HTTP client. Everything here is live on flow.finalpos.com with no configuration changes on your side. Connect your MCP client and build. Setup details are at finalpos.com/mcp.