Built for focus. Compose blocks, write prompts, swap models, run real input, read the trace back — every tool you need in one workspace.
Noukai gives you a dedicated workspace for AI engineering. Every surface — the prompt editor, the block navigator, the run viewer, the evals — is shaped for the pipeline you're building.
The prompt editor takes the main space in every block. You shape the prompt, watch the schema fold around it, and run it against a sample input — without leaving the surface.
Read the user message {{input.message}} and tag the intent. Return JSON: { intent: 'billing' | 'technical' | 'other', confidence: 0..1 }
When the prompt is right but the model isn't, swap it. Every block carries its own model — try a different provider, compare runs, never touch the wiring.
While you build, the pipeline you're shaping stays in front of you. Every panel and tool wraps around it — your attention on the target stays sharp.
The block navigator maps out every pipeline in your project — every lane, every block, how they connect. As the project grows, the shape stays clear.
When the pipeline is ready, wire it in. The Noukai SDK lives in TypeScript and Python — one client, one execute call, your pipeline runs from your app.
import { Noukai } from "@noukai/sdk"; const noukai = new Noukai({ org: "acme", project: "support", }); const result = await noukai .flow("classify") // pipeline name .execute({ message });
Noukai ships an MCP server that exposes every pipeline operation as a tool. Wire Claude Code, Cursor, or any MCP-aware editor into your workspace — they read your flow, write the next block, and run the trace without ever leaving their window.
WHAT YOUR EDITOR DRIVES
Sketch lanes and blocks from your goal — plain English in, pipeline shape out.
Write the prompts, pick the models, wire the schema on every block.
Run the pipeline against real inputs, read the trace back, patch what slips.
Tune prompts against the cases that matter — automatically, in the background.
When the next step isn't an LLM call — when it's a calculation, a sub-pipeline you've already built, a loop over rows, or a tool the model needs to reach for — Noukai already has the block. The LLM stays where only it makes sense.
Drop in the block that fits the step. Code for the deterministic parts, sub-pipelines for the composable ones, loops for the rows, passthroughs for the routing — and a real LLM block where the model earns its keep. More block types are on the way.
Every block carries its schema. Declare the fields you expect in, the fields you return out — and when you chain the next block on, let it inherit the shape so the data flows through without glue code in between.
Declare tools in your server, hand them to the pipeline when you call it. LLM blocks pick them up at run time — the model reaches for them when the prompt asks, you read every call back in the trace.
tool calls
Save a checkpoint when you land a prompt that earns it. Pick which one runs in production, swap back the moment the new version slips, diff any two — your prompts versioned the way your code is.