Skip to main content
BuildHow-to· Updated September 14, 2026· Jackson Mclean

Build a flow with your own AI tool over MCP

To build a flow with your own AI tool over MCP, connect the tool to Flow, describe the checkout you want in plain language, let the tool load Final's current scaffold, then ask it to validate, preview and deploy. Your tool calls the right Final tools in the right order; you review warnings, test the preview, and switch to the new flow from Station Home.

Once your AI tool is connected to Flow over MCP, you build a flow with your own AI tool the same way every time: describe what you want, let your tool load Final's current scaffold, validate, preview, deploy. You talk to your tool in plain language and it calls the right Final tools in the right order. This article covers what to say at each stage and what to look for before you put a flow on a live Station.

Before you start

You need:

  • A Final account.

  • An AI tool that supports MCP, such as Claude Code, Cursor, Codex, ChatGPT or Gemini CLI.

  • Your tool connected to Flow. On the Flow home, type a prompt and choose Connect your own AI (MCP), then paste the block Flow gives you into your tool. If you already have a flow open in Flow Studio, use the cable icon in the chat rail and choose Connect your AI tool. The article "Connect your own AI to Flow Studio (MCP)" walks through this step by step.

Step 1: Describe the checkout you want

Tell your tool what the flow needs to do, in plain language. The more specific you are about how your business runs, the less you fix later. Cover:

  • Service type: table service, counter service, self-checkout or kiosk.

  • Key interactions: split bills, modifiers, tip collection, a customer-facing display.

  • Hardware: tablet, phone, countertop terminal or kiosk screen.

Example prompt:

Build a counter-service flow for a café. It needs a product grid with categories, a cart with quantity controls, a tip screen with preset percentages, and a tap-to-pay step. Station Home should show end-of-day totals.

You do not need to name Final tools or tell your AI what order to work in. The Final MCP server gives it those instructions when it connects.

Step 2: Let your tool load the current scaffold

Before it writes anything, your tool fetches Final's current flow scaffold: the starting files and the version of Final's command framework your flow has to match. This grounds everything it generates in what Final runs today, and it is the single biggest predictor of a clean first result.

This normally happens on its own. If your tool jumps straight into writing code, say: "Load the Final scaffold before you build."

Step 3: Know what the validator checks

Three mistakes show up most often in first drafts. Validation catches all three, but knowing what they are helps you understand what your tool is fixing.

  • Prices or products typed into the flow. Prices, product names and SKUs should come from your catalog, so a change in Merchant Hub shows up at the till without rebuilding the flow.

  • Totals calculated inside the flow. Final calculates totals, tax and tips on the server. Math inside the app leads to rounding and reporting problems on a live till.

  • No way back to Station Home. Every flow needs a Station Home button so staff can always get back to the start.

Step 4: Validate before you preview

Say: "Validate the flow." Your tool runs Final's checks and reports back. You will see three things:

  • Warnings. The three issues above, plus one more: a tip added as a cart fee instead of using Final's tip field. Warnings do not block deployment, but treat the catalog and money ones as blockers. They cause real problems in production.

  • Type check. Whether Final's code check ran. If it reports unavailable, the check was skipped, so ask your tool to run its own type check before moving on.

  • Version match. Whether your flow still matches the current Final framework. If your tool reports a mismatch, say "Reload the Final scaffold and update the flow," then validate again.

Each warning comes with repair context your tool can act on directly. Ask it to fix every warning, then validate again. Repeat until the result is clean.

Step 5: Preview the flow

Say: "Start a preview." Your tool opens the flow in Final's preview environment and gives you a link. Test the interactions that matter for your business: navigation, modifier selection, the tip screen, getting back to Station Home.

The preview is a sandbox, not a live till:

  • Payments always succeed instantly. No real charge is made.

  • Totals are sample values, not live card totals.

  • Some cart and filtering behaviour is simplified.

Your tool receives this same list when the preview starts, so it knows not to chase "bugs" that are preview limits. If something does not match what you asked for, tell your tool, let it change the flow, and validate again before you deploy.

Step 6: Deploy

Say: "Deploy the flow." Deployment stops if the flow contains a password or API key typed directly into the code. Flows can be shared with other organizations, so Final never lets a secret ship inside one. Ask your tool to store the credential in Final's secrets vault instead, then deploy again.

Once deployed, the new version is available to your Stations right away. Open Station Home on a Station and switch to the new flow or version.

Tips

  • Never skip the scaffold. A flow built without it may reference parts of Final that have changed.

  • Validate after every round of edits. Each correction your tool makes can introduce something new.

  • Treat catalog and money warnings as blockers, even though the validator calls them advisory.

  • Want to finish by hand? Ask your tool to open the flow in Flow Studio, and carry on there with preview and deploy.

For developers

If you drive the tools directly, the steps above map to get_scaffold, validate_flow, start_preview, sync_preview, deploy_flow and open_in_builder. The full tool catalogue, transport details and client configuration are at finalpos.com/mcp.

Step by step

  1. 1

    Describe the checkout you want

    Tell your AI tool the service type, key interactions and hardware in plain language. You do not need to name Final tools; the MCP server handles the order.

  2. 2

    Let your tool load the scaffold

    Your tool fetches Final's current flow scaffold before writing code. If it skips this, say: Load the Final scaffold before you build.

  3. 3

    Validate the flow

    Say: Validate the flow. Review the warnings, type check and version match, ask your tool to fix every warning, and validate again until clean.

  4. 4

    Preview the flow

    Say: Start a preview. Test navigation, modifiers, the tip screen and Station Home. Payments and totals in the preview are simulated.

  5. 5

    Deploy the flow

    Say: Deploy the flow. If deployment stops on a typed-in secret, move it to Final's secrets vault and deploy again.

  6. 6

    Switch to the new flow on your Station

    Open Station Home on a Station and switch to the new flow or version. It is available right away.

Frequently asked questions

Do I need to know the Final tool names to build over MCP?

No. The Final MCP server gives your AI tool the instructions and the order of operations when it connects. You describe what you want in plain language and review what comes back.

What does it mean when validation says the type check was unavailable?

Final's own code check did not run on that pass. Ask your tool to run its own type check before you move on to preview, so type errors are not missed.

Why did deployment stop with a credentials error?

A password or API key was typed directly into the flow's code. Flows can be shared with other organizations, so Final blocks that. Ask your tool to store the credential in Final's secrets vault and deploy again.

Do warnings stop me from deploying?

No, they are advisory. Treat the catalog-data and money-math warnings as blockers anyway; they cause real problems on a live till.

Can I finish the flow in Flow Studio instead of my AI tool?

Yes. Ask your tool to open the flow in Flow Studio and continue there with preview and deploy. Everything it built is already in your account.

What does the preview not simulate?

Payments always succeed instantly with no real charge, totals are sample values rather than live card totals, and some cart and filtering behaviour is simplified. Your tool receives the same list when the preview starts.