Introduction

SWEny turns natural language into reliable AI workflows. Describe what you want done, and SWEny builds a DAG — a directed acyclic graph of nodes, each with a focused instruction, scoped tools, and structured output. Then it runs it, tracking every node, tool call, and routing decision.

sweny workflow create "research competitors, gather pricing and features, \
  synthesize a comparison, and produce an executive brief"

One sentence in, a full workflow out. Refine it with natural language, run it immediately, or deploy it to GitHub Actions.

Why a DAG?

A single prompt that says “research competitors and write a report” works sometimes and fails unpredictably. SWEny splits that work into discrete nodes with explicit control flow:

• Reliable — each node has a focused instruction and scoped tools. If a node fails, you know exactly where and why.
• Observable — the executor emits structured events (node:enter, tool:call, node:exit, route) so you can log, trace, and audit every step.
• Repeatable — the same workflow definition produces the same execution path every time, regardless of who triggers it or where it runs.

How it works

You don’t need to write YAML. sweny workflow create generates workflows from plain English, and sweny workflow edit modifies them. But here’s what’s happening under the hood:

  ┌─────────────┐
  │  Research    │  ← entry node
  │  Competitors │    skills: github
  └──────┬───────┘
         │
  ┌──────▼───────┐
  │  Gather      │    skills: github
  │  Data        │
  └──────┬───────┘
         │
  ┌──────▼───────┐
  │  Synthesize  │    (no external tools needed)
  └──┬───────┬───┘
     │       │
  when:    when:
  data     zero
  found    results
     │       │
  ┌──▼──────┐ ┌▼──────────┐
  │ Write   │ │ Flag Data  │
  │ Brief   │ │ Gap        │
  └─────────┘ └────────────┘

At each node, Claude receives the instruction, the tools from the node’s skills, and context from all prior nodes. After a node completes, the executor evaluates edge conditions — written in natural language — to decide which node runs next.

The executor uses headless Claude Code as the LLM backend. MCP servers for GitHub, Linear, Sentry, Datadog, and others are auto-injected based on your configuration.

Three ways to use it

Method	What it does
CLI	Build and run workflows from your terminal. The primary way to create workflows and get things done.
GitHub Action	Deploy workflows to CI for scheduled automation. Built-in triage monitors production errors.
Studio	Visual DAG editor and live execution monitor. Watch workflows run node-by-node.

Looking for inspiration? Browse the Workflow Marketplace for ready-to-run workflows you can copy, customize, and run.

What people build with it

SWEny works for anything you can describe as a sequence of steps:

• Content generation — generate blog posts, run them through LLM quality judges, publish passing content (used for kidmath.ai)
• Security audits — scan commits for secrets, review PRs, check dependencies, file tickets for findings
• Competitive analysis — research competitors, gather data, synthesize reports, create action items
• Production triage — monitor errors, investigate root causes, create issues, open fix PRs (built-in workflow)
• Product launch prep — research launches, draft copy with quality gates, create checklists

Built-in workflows

SWEny ships two production-ready workflows:

Triage — monitors your observability platform for errors, performs root cause analysis, creates an issue, and notifies your team. Runs on a schedule.

Implement — takes an existing issue, analyzes the code, writes a fix, opens a PR, and notifies. Triggered manually or chained from Triage.

You can also build custom workflows from natural language, YAML, or with the Studio visual editor.

Next steps

• Quick Start — install and create your first workflow in under a minute
• Core Concepts — understand workflows, nodes, edges, and skills in depth
• CLI Examples — real-world workflow examples from one-liners to complex pipelines
• End-to-End Walkthrough — follow a real triage run from error spike to fix PR