Local-first architecture docs
Document AI-agent architectures, visually.
A local-first visual editor that gives multi-agent systems a shared, inspectable map: typed components, semantic relationships, and the behavior behind them. Documentation, not a runtime.
- No account
- No backend
- No API key
- Open source (MIT)
Features
Structure on the canvas. Behavior in the inspector.
- Orchestrator
- Agent
- Tool
- Memory / Data Store
- External System
- Decision / Routing Rule
Model
- Six component types
- Agent-specific building blocks, each with fields that fit its role.
- Ten semantic relationships
- Delegates To, Supervises, Uses, Reads From, Writes To, Shares Context With, Reports To, Hands Off To, Depends On and Generic.
- Groups
- Named, non-executable containers. Moving a group moves its members.
Canvas
- Quick Find
- Ctrl/Cmd+K or / searches by name, type or group.
- Arrange
- Align, distribute and tidy. Geometry only; relationships stay intact.
- Component colors
- Color one component or a selection. Presentation, not semantics.
- Undo, redo, duplicate
- And safe delete, with a confirmation that explains what goes.
Document
- Type-aware inspector
- Purpose, responsibilities, boundaries, delegation and failure handling.
- Markdown notes
- Edit/Preview toggle. Raw HTML and remote images are never rendered.
- Autosave
- Saves locally with a visible save state and a retry if a save fails.
- Local versions
- Automatic, manual and pre-restore snapshots you can restore.
Export
- JSON
- Canonical files for backup, review and re-import.
- Markdown
- A deterministic, human-readable review document.
- PNG
- The whole diagram as an image for reviews and issues.
- Validated import
- Checked before saving, always a new architecture; v1 files migrate.
How it works
From blank canvas to shareable artifact.
-
Model
Add an orchestrator, agents, tools, memory, external systems and decisions. Connect them with typed relationships.
-
Document
Describe purpose, behavior, boundaries and failure handling in the inspector. Group, find and arrange to stay readable.
-
Export and share
JSON for a durable backup, Markdown for a review document, PNG for a visual share.
Who it's for
For people who need to explain how an agent system works.
- AI engineers Map agents, tools and memory before, during and after implementation.
- Software architects Review designs in a framework-agnostic vocabulary with portable artifacts.
- Researchers Describe multi-agent designs with structure and behavior, not just prose.
Scope
A documentation tool, on purpose.
It is
- A visual documentation tool for AI-agent architectures
- A framework-agnostic vocabulary for components and relationships
- A structured place for behavior, notes and design rationale
- A static web app with open JSON, Markdown and PNG artifacts
It is not
- An agent runtime, orchestration framework or workflow engine
- An LLM client, prompt runner or model gateway
- A simulator, observability platform or debugging tool
- A cloud service, account system or real-time collaboration product
Relationships are semantic documentation: arrows describe how components relate, not execution order or data flow.
Your data
Stays in your browser.
- Local-first. Architectures live in IndexedDB for this browser origin. Autosave writes locally.
- Nothing phones home. No account, no application backend, no analytics or telemetry, no runtime AI calls.
- JSON is the source of truth. Use it to move an architecture to another browser or device.
Start mapping your agent system.
It runs in your browser. Nothing to install, nothing to sign up for.