Workflow · Real tool

AI Workflow Builder

Chain LLM steps. Each step's output is available to the next via {{stepId.output}}; initial values come from {{input.<key>}}. Pick a preset, edit any prompt, watch the chain execute step by step with per-step latency.

Pick a workflow

Library →

v0.4 adds parallel: true — consecutive parallel-flagged steps fan out in one batch via Promise.allSettled. Combined with v0.3's runIf conditional execution you get both fan-out and branching. Step types: llm · fetch · assert. Outputs flow via {{stepId.output}}.

Classic content pipeline. A researcher lists sub-topics; a summariser turns them into two paragraphs; a closer distills it into a single tweet.

Input {{input.topic}} — Topic

Flow diagramauto-rendered

flowchart LR
  research["Research"]
  summary["Summarise"]
  tweet["Tweet"]
  research --> summary
  summary --> tweet
  classDef pll fill:#f5f3ff,stroke:#8b5cf6,color:#5b21b6,stroke-width:1.5px
  classDef cnd fill:#fffbeb,stroke:#d97706,color:#92400e,stroke-width:1.5px

Research

llmIdle

Output

Idle.

Summarise

llmIdle

Output

Idle.

llmIdle

Output

Idle.

How this fits the AI-ops toolkit

ai-eval evaluates a single LLM call against assertions. Workflow Builder chains many calls and feeds their outputs into each other.
Multi-Agent PR Reviewer is a fixed 4+1 workflow (four parallel reviewers → one aggregator). Workflow Builder is the user-defined version.
Multi-Agent Orchestrator is the educational demo of the same pattern; this is the workable tool.

v0.1 limits: serial steps only (parallel + branching on the roadmap), one provider (OpenRouter via /api/demo/generate), each step capped at 400 output tokens. Future: branching, parallel forks, fetch / assert step types, save+share workflow URLs, export to ai-eval YAML.