“We didn’t just build Noukai — we dogfood it on Nouko, so every prompt, pipeline, and fix lands the way our users would want it.”
abigcircle
Language learning, in production on Noukai
脳海 · Sea of the mind
Noukai is the workspace for AI engineers building scalable pipelines. Compose lanes of blocks, write prompts, swap models, run real input, read the trace back — every tool you need, shaped for the work.
v0.1 · public betaChangelog
Processing Pipeline/Transform
processing-pipeline
Processing Pipeline
Data Ingestion
Fetch Source
Validate Schema
Processing Pipeline
Transform
Enrich
Classify
Merge Results
Output
No checkpoint
152 charstokens: 4096 · temp: 0.5
一／ Build
Lanes, not infinite canvas. Pipelines read left-to-right in lanes — like a score, not a circuit board.
Data Ingestionv3
Processing Pipelinev7
Output
PIPELINE EDITOR
The visual UI for prompt engineering — without the spaghetti.
Every flow is a stack of lanes. Each lane runs its blocks left-to-right, passing outputs forward. Parallel branches sit in tinted brackets — visually grouped so you always know what runs together.
- Drag to reorder lanes and blocks
- Branching with v-split brackets, no graph nodes
- Passthrough blocks wire outputs without prompts
- Rename, type-swap, and remove inline
二／ Debug
Debug every step. Inputs, outputs, latency, cost — in one trace, in the order they happened.
DEBUG & OBSERVABILITY
When something’s off, you’ll know exactly where.
Every run produces a full step-level trace with inputs, outputs, token counts, and latency. Compare runs side-by-side, replay any step with edited inputs, and see exactly how your pipeline spends its budget.
- Step-level I/O inspection for every run
- Latency bars show where time is spent
- Diff inputs between passing and failing runs
- Replay any step with modified inputs
run · 26m02k · 2.7s · $0.00411 step failed
01
Classify intent
412ms02
Lookup customer
78ms03
Plan response
890ms04
Draft answer
1262ms05
Validate schema
41msTotal latency
2.7 s
Total cost
$0.0041
三／ Ship
Pipelines you can trust in production. Versioned, promotable, rollback-able. Like code, because it is.
Draft2 unpublished
Unpublished changes(2)
4 versions
v4production
May 2 · 10:00tighten validator on edge cases
v3
May 1 · 14:00switch draft step to gpt-5
v2
Apr 29 · 11:00add language fork, retire fallback
v1
Apr 25 · 09:00PUBLISHING & VERSION CONTROL
Ship with confidence. Roll back in seconds.
Every published pipeline version is immutable and tagged. Promote from draft to staging to live with a single action. When something goes wrong, roll back to any previous version instantly — no redeploys, no downtime.
- Immutable versioned snapshots of every lane
- Draft → staging → live promotion workflow
- One-click rollback to any prior version
- Full diff between any two versions
四／ Call
Every pipeline is a URL. A stable URL, a bearer token, the version of your choice. Drop it into any stack you ship.
API ENDPOINTS
Call from anywhere. Versioned by default.
Publish a pipeline and it’s already an endpoint — slug-addressed, scoped to your organization. Pass an API key, call it from any language. Pin a version or follow latest published — the contract holds while the pipeline evolves underneath.
- Per-pipeline URL — stable and slug-addressed
- Bearer-token API keys, scoped per organization
- Pin a version, or follow latest published
- Stream with SSE, or one-shot JSON
Request
POST
noukai.xyz/api/v1/acmeorg/support-botproject/triagepipeline/execute
Headers
AuthorizationBearer nk_live_••••••••••••
curl
$ curl -X POST https://noukai.xyz/api/v1/acme/support-bot/triage/execute \
-H "Authorization: Bearer $NOUKAI_API_KEY" \
-d '{ "input": "..." }'200 OKv41.2s$0.0041
五／ System
The system absorbs complexity. You never wear it.
05.1
One-click model switching
Swap any block to a different model without touching prompt text. Compare outputs across providers in a single run.
05.2
Structured output, by default
Define a JSON schema on any block. Noukai enforces it — valid output on every run, no parsing glue code.
05.3
Template version control
Every prompt edit creates a new version. Diff any two, restore any prior state, see who changed what and when.
1tighten validator output formatApr 30
2add context window guardApr 29
3swap model to claude-sonnet-4.5Apr 27
4add metadata extraction stepApr 25
5initial classify-intent promptApr 23
六／ In practice
Teams shipping with Noukai. From first pipeline to production traffic.
If you’re shipping with Noukai, let us know! We’d love to feature you here!