Stepwise

Step into your flow.

Build multi-step AI workflows visually or in YAML. Mix shell scripts, LLM calls, autonomous agents, and human approval gates — with automatic parallelization, cost tracking, and a real-time web UI.

generate-homepage.flow.yaml — a real 10-step flow

Author flows visually

A full authoring environment built into Stepwise. Write YAML in the editor, see the DAG update live, or let AI generate flows for you.

Visual Flow Editor

YAML + live DAG, side by side

Editor screenshot
CodeMirror 6 YAML Live DAG preview Click-to-edit nodes Add/delete steps Registry browser

AI-Assisted Authoring

Describe what you want, get working YAML

AI chat screenshot
Streaming chat One-click Apply Context-aware Quick actions YAML generation

The engine

A portable YAML workflow engine with the features you need and nothing you don't.

YAML-first simplicity

Flows are portable YAML files — or directories with co-located scripts and prompts. No DSL, no vendor lock-in. Automatic parallelization from dependency analysis.

Mixed executor types

Shell scripts, LLM calls via OpenRouter, autonomous agents with ACP, and human approval gates — all in a single flow. Each step uses the right tool.

Route steps & control flow

Conditional branching with route steps, expression-based exit rules, loop-back on failure, for-each parallel fan-out. First-match dispatch with inline or registry flows.

Context chains

Multi-step agent workflows with session continuity. Prior conversations compile into structured context blocks so each agent builds on the last.

Flow registry

Share, discover, and install community flows. Publish with stepwise share, install with stepwise get, or browse the registry from within the editor.

Cost & time controls

Cap spend per LLM call, set timeouts and retries per step, limit loop iterations. Decorators for timeout, retry, fallback, and notification — composable per step.

Define. Run. Watch.

A real flow in 30 lines of YAML.

content-pipeline.flow.yaml
name: content-pipeline
steps:
  research:
    executor: llm
    model: anthropic/claude-sonnet-4
    prompt: "Research trends in {topic}."
    outputs: [insights, sources]
    inputs:
      topic: $job.topic

  draft:
    run: python scripts/draft_article.py
    outputs: [content, word_count]
    inputs:
      insights: research.insights

  review:
    executor: human
    prompt: "Review this draft."
    outputs: [decision, feedback]
    exits:
      - when: "outputs.decision == 'approve'"
        action: advance
      - when: "attempt < 3"
        action: loop
        target: draft

  publish:
    run: python scripts/publish.py
terminal
$ stepwise run content-pipeline --watch

 research        (2.3s) — 847 tokens, $0.08
 draft           (1.1s) — 1,247 words
 review          Waiting for human input...
  → Decision: approve
 publish         (0.8s) — /articles/ai-workflows

 Completed in 4.2s | 4 steps | $0.08 total

Start in 60 seconds

Three commands to your first flow.

1

Install Stepwise

$ curl -fsSL https://raw.githubusercontent.com/zackham/stepwise/master/install.sh | sh
2

Create and run a flow

$ stepwise init && stepwise new my-flow
$ stepwise run my-flow --watch

Creates a .stepwise/ project, scaffolds a flow directory, and runs it with the real-time web UI.

3

Open the editor

$ stepwise serve

Launches a persistent server with the visual flow editor, DAG visualization, AI chat, and the registry browser — all in your browser.

See it in action.

A real 10-step flow that ships a product:

Scripts ingest a codebase. LLMs plan and generate. For-each fans out to produce sections and pages in parallel. Human review gates catch issues. Everything runs on automatic — you just watch.

Browse the flow gallery
ingest-repo (0.8s) — parsed README, docs, 47 source files
plan-design (12.4s) — design tokens, component specs
gen-mockup (8.2s) — Imagen 4, 1920x1080
code-hero (15.7s) — Gemini multimodal, image to code
gen-copy (6.1s) — 2,847 words of marketing copy
generate-sections (22.3s) — 6 parallel, review + loop
generate-pages (31.9s) — 4 parallel, review + loop
assemble (3.2s) — merged sections, nav, footer
write-output (0.4s) — 8 files, 127kb
Completed in 47.2s | 10 steps | $4.73