StepbookStorybook for typed pipelines

Edit one step, re-run only what changed, then diff the result to see exactly what your edit did — locally, in your repo. Especially for LLM pipelines.

Get Started

Why Stepbook

View on GitHub

Change one step, not ten

The cache is keyed by each step call's input and its source-file hash. Edit a prompt or parser and only that step — plus any step whose input changed as a result — re-runs. The rest is served instantly from cache. That's half the loop.

See exactly what changed

Diff any two runs structurally — field-level output drift and assertion flips, instead of eyeballing two blobs of JSON. Cheap re-runs and structural diffs are the loop stepbook is built around — that's the other half.

Typed, asserted, snapshotted

Each step is a typed async function with an optional Zod schema and assertions, composed in plain TypeScript — a pipeline calls steps in a line, an agent in a loop. Every call is a first-class node in the run's trajectory, so a failure points at exactly where it broke.

Built for the LLM dev loop

Stability clustering, outlier detection, and cost & token telemetry — the tooling that's most missing when your steps are non-deterministic.

A CLI humans and agents share

Every command emits a versioned --json envelope with stable exit codes, so AI coding agents can drive stepbook as a tool — not just sit beside it.

Local-first, no SaaS

Everything lives in a state/ directory in your repo — JSONL run store, file-based snapshot cache, run blobs. Your data never leaves your machine.

A dashboard when you want one

A React + Vite inspector with History, Compare, Checks, Telemetry, Stability, and Query panels. Optional — the CLI does everything headless.

Stepbook dashboard — pipeline tree on the left, the selected step's output in the canvas, checks in the bottom drawer.

The dashboard inspecting the bundled examples/demo pipeline.