Agent Blueprint Engine — AI skill for Claude Code

Agent Blueprint

**Design once. Test it. Compile it everywhere.**

Agent Blueprint is a visual IDE for designing, validating and compiling AI agent systems. You describe agents, skills, workflows, iron laws, rules, hooks, gates, tools, references and memory once, as a _Blueprint_; the compiler turns that Blueprint into the files each AI coding harness expects, and the validator tells you what is missing, contradictory or unsupported before you ship it. Bring your own OpenAI-compatible endpoint and a model drafts, improves and reviews the Blueprint with you — every proposal a change you read before it is applied. Connect GitHub and the finished repository is one push away, previewed before anything is written, and a repository's `blueprint/` opens back as a project.

Everything happens in the browser. There is no server, no account and no database: a project lives in IndexedDB, in a folder on your disk, in a ZIP, or in a Git repository, and the two optional route handlers are stateless relays that exist only to work around browser limits — one for AI endpoints that refuse browser calls, one for GitHub sign-in. Keys and tokens stay in the browser and never reach a Blueprint, an export, a push or a log.


What it does

**Design.** Every artifact has a visual form and a source tab showing the project file itself, with a Markdown preview. Workflows are edited as the drawing they are. Ten starter Blueprints and 36 artifact templates give you something to open rather than an empty editor, a command palette (`⌘K`) creates, navigates and validates, and `/tutorial` walks the whole loop and then every kind of artifact — what it is, how to add one, and what each harness gets.

**Validate.** Structural, semantic, orphan, contradiction and requirement rules, each with a stable code and a navigable reference. The health bar opens the findings behind every count.

![Evaluation](apps/web/src/assets/screenshots/eval