Config, not code
Your AI defines entities, Logic Blocks and crews as short configs. The platform turns them into a running API.
Why token-minimal
Spend tokens on decisions, not on rewriting code.
AI app builders keep a model in the loop for every change, and you pay for it in credits. nostackai uses your AI to design the backend, then runs it as plain, deterministic infrastructure. Tokens go only to the steps that need judgement.
{
"entityName": "order",
"UpdateParams": [
{ "name": "customer", "colType": "lookup" },
{ "name": "total", "colType": "number" },
{ "name": "status", "colType": "string" }
],
"baseKeys": { "pkTemplate": "order#{id}" }
}This call is the entire definition. The platform generates the endpoints, search, access checks and API docs from it.
Where the tokens go
Five places an AI backend spends tokens. We cut four.
| Where tokens get spent | Typical credit-based AI builder | nostackai |
|---|---|---|
| Building the first version | The agent writes and runs a full code base, then fixes what fails. | Your AI makes short config calls the platform validates on the spot. |
| Changing it later | Each prompt re-reads files, regenerates code and re-tests. Prompt 50 costs more than prompt 1 as the code grows. | A change merges a few fields into one definition. Prompt 50 is about the size of prompt 1. |
| Knowing what is possible | The agent explores, guesses and retries. | get_platform_capabilities and a live API reference tell it exactly what exists. |
| Running in production | Whatever the generated code does. | No model in the request path, except the AI steps you add. |
| AI inside the app | Usually platform credits, or keys wired in by hand. | Your keys on every plan, per workspace, with tokens and cost shown for each workspace. |
A comparison of how each approach works, not a measured benchmark.
How it stays small
Five rules the platform is built on.
No guessing
The MCP server tells the AI which field types, step types and connectors exist, so it gets it right the first time.
No model in the path
Requests run as ordinary serverless code. A thousand orders cost a thousand API calls, not a thousand prompts.
Small when you use it
AI steps and crews get only the fields you map, answer in typed shapes, and run under token caps you set.
No credit markup
AI steps use your provider keys on every plan, per workspace if you like. The usage page shows tokens and cost for each workspace.
When your app does use AI
AI steps stay small too.
Many apps need AI at run time: an ai_call step in a Logic Block, or an agent crew. nostackai keeps those calls short and bounded, and runs everything around them without a model.
- Only what you map. An AI step receives the fields you pass it, not a conversation, a codebase or a whole record set.
- Typed answers. Ask for a category, a yes/no, a score or JSON fields. Short answers, no re-prompting to fix the format.
- A model per step. A small model for classifying, a larger one only where the task needs it.
- Caps you set. A token limit per step. Crews add a token cap per run, a turn cap and a tool-call cap per agent.
- Guardrails before the model. Blocked or off-topic input stops a turn before any tokens are spent.
- Memory that stays small. Agents keep a sliding window or a summary instead of replaying the full history.
01get_entityticket · {{ input.id }}no AI
02ai_callclassify · billing / bug / other1 short call
03conditioncategory = billingno AI
04update_entityticket · queue, priorityno AI
05connectorslack · #billingno AI
One AI step with a one-word answer. The other four steps run as ordinary code.
FAQ
Common questions
Does nostackai call an AI model when my API runs?
Only in the steps where you put one: an AI step in a Logic Block, or a crew. Reading and writing data, search, access checks, events, schedules and flags run without a model.
Which AI builds the backend?
Yours. You connect Claude Code, Claude Desktop, Cursor or another MCP client, and it uses the subscription or API key you already have. nostackai does not resell AI credits.
Who pays for the AI steps inside my app?
You do, at your provider's price, with your own keys, on every plan. Each customer workspace can use its own keys, so AI cost can be passed through to the customer who incurs it.
Why does config use fewer tokens than generated code?
A config call is short and the platform validates it. A change merges a few fields into an existing definition. There is no file tree for the AI to read again, regenerate and debug on every request.
What if config is not enough?
Write a Glyph: a small Python or Node function that runs next to your data. You write code only where it is needed.
Can I see what the AI steps cost?
Yes. The usage page breaks AI input tokens, output tokens and cost down by customer workspace, and every crew run reports its own tokens and estimated cost.
Can I use a model you do not list?
Yes. An org admin can define any provider with an OpenAI-compatible endpoint, and your Logic Blocks and agents can use it like the built-in ones.
Ready when you are
Describe it once. Run it forever.
Connect your AI client, build your backend, ship to production.