aiworkbench

A native macOS desktop app for orchestrating Claude Code agents across a visual task canvas.

Work in progress — rough edges expected.

Workflow

1. Create a project and point it at a local directory
2. Add tasks on the canvas — leaf tasks for single prompts, container tasks for grouped subtasks
3. Draw dependency edges to enforce ordering between tasks
4. Queue tasks individually or all at once via the "+ all" button
5. Hit play — the runner launches Claude Code in a PTY session for each task in dependency order
6. Watch live status (queued → running → done) in the queue panel and reflected on the canvas nodes
7. Claude agents can read and update task state via the embedded MCP server written to each project's .mcp.json

Concepts

• Leaf task — a single unit of work with a prompt, run by one Claude Code session
• Container task — a group of ordered subtasks, each run sequentially by Claude Code
• Dependencies — draw edges on the canvas to enforce ordering; queuing a task auto-queues its deps first, removing a task cascades to remove dependents
• Queue — ordered list of pending/running/done items; a runner per project processes them one at a time
• MCP server — HTTP JSON-RPC server started at launch; Claude agents connect via .mcp.json written to each project directory

Requirements

• macOS
• Wails v2 (go install github.com/wailsapp/wails/v2/cmd/wails@latest)
• Go 1.21+
• Node 18+
• Claude Code installed and on $PATH

Development

wails dev

Hot-reloads the frontend. Go changes require a restart.

Build

wails build

Produces a .app bundle in build/bin/.

Data

The SQLite database lives at ~/Library/Application Support/aiworkbench/aiworkbench.db. Schema migrations run automatically on startup.