Skip to main content
T1 starter · local-first: the whole loop on your machine — a real HTTP fetch, a local model, a verifiable trace. No API key exists in this tutorial, so there is nothing to leak and nothing to meter.

​
The job

A morning brief from a live source, every day, the same way. One file captures it; your laptop runs it; the trace proves what happened.

​
Prerequisites

  • nika installed: brew install supernovae-st/tap/nika (other paths)
  • Ollama running with a small model: ollama pull qwen3.5:4b (any local model works — swap the model: line)

​
The file

Save as daily-brief.nika.yaml: 
nika: v1
workflow: daily-brief
model: ollama/qwen3.5:4b

tasks:
  - id: fetch
    invoke:
      tool: "nika:fetch"
      args: { url: "https://hn.algolia.com/api/v1/search?tags=front_page" }

  - id: brief
    depends_on: [fetch]
    infer:
      max_tokens: 300
      prompt: |
        Five bullet points, most signal first: ${{ tasks.fetch.output }}

outputs:
  brief: ${{ tasks.brief.output }}
Swap the URL for any JSON source — an RSS-to-JSON bridge, your issue tracker’s API, a status page.

​
Check, then run

nika check daily-brief.nika.yaml
The static audit passes before a single token is spent — plan, types, tools, args, secret flows. One honest wrinkle to read, not fear: 
⚠  COST   $0.0000 – $0.0000 FLOOR (unbounded tasks present)
   brief  ollama/qwen3.5:4b  UNBOUNDED — no catalog price (local/unknown model)
A local model has no list rate, so the engine says unbounded instead of pretending $0.00 is a promise. Local work is never blocked and never metered — the flag exists so a paid uncataloged model can’t hide behind a fake zero (cost honesty). Then: 
nika run daily-brief.nika.yaml
A real run of this exact file (Apple Silicon, qwen3.5:4b): 
✔  fetch  invoke · nika:fetch        762ms
✔  brief  infer · ollama/qwen3.5:4b  48.0s
── 2/2 done · ≥ $0.00 (1 unpriced) · elapsed 48.7s ─────────────
  trace: .nika/traces/2026-07-09T22-54-50Z-e843.ndjson · 8 events · chain 394719d2…

​
The receipt

nika trace verify .nika/traces/<run>.ndjson   # path from the run card's trace: line

OK — 8 events · chain intact · head 394719d24361e1a6…
The trace is a hash chain: edit any event after the fact and verify exits non-zero. Your brief comes with proof of what ran, in what order, against which inputs — even offline.

​
Make it daily

Cron it, or wire the GitHub Action if the file lives in a repo — every edit to the workflow gets a static check before it lands. When you want the run itself in CI, that stays a deliberate step in your own job, with your own budget flags.

​
Where to go next

  • Add a permits: boundary — nika check --infer-permits writes the tightest one, and the workflow becomes default-deny (security).
  • Persist the brief: a third task with nika:write and a content: binding (starters show the shape).
  • Mix providers: keep fetch local, point brief at a cloud model with --max-cost-usd — the budget gate prices the file’s own model before anything starts (cost honesty).