AI Kanban

Local-first human–AI work

Give every piece of work a place to think.

A calm board where plans, decisions, agent turns, research memory, and finished results stay together—in a folder you control.

Personal research

Choose our July family vacation

3 optionsNeeds your view
Evidence gathered12 sources · current
AI updatedTradeoffs are ready to review
No backend Folder-backed SQLite Works without an AI provider

Shape the work

New card

Card workspace

Find the right work

Board filters

Cooperative writer baton

Agent control

1
Start your agent

Open Copilot CLI, Claude Code, or a compatible agent in the Board Workspace folder.

2
Let it read the skill

The root instructions point to one canonical, versioned skill bundle.

3
Grant control when ready

The browser saves and closes its writer before an approved agent can mutate SQLite.

Fallback instructions and diagnostics

Read manifest.json and .agents/skills/ai-kanban/SKILL.md, then run the skill preflight.


        

        
      

    

    
      

        

          
Provider-neutral boundary
AI handoff

          
        

        
        

          
          
          
        

        

          
Compact intent, context, plan, activity, and requested response shape. Nothing is uploaded.

          

          
          

            
            
          

        

        
        
      

    

    
      

        

          

          
        

        

        
      

    

    
      

        

          
External change detected
Your draft is safe.

          
        

        
AI Kanban stopped before overwriting the folder. Compare revisions, reload the durable workspace, export your draft, or explicitly recover.

        

        

          
          
          
        

      

    

    
      

        

          
Local data boundary
About workspace

          
        

        

        

          Technical model
          
board.sqlite is canonical. A dedicated Worker owns the browser connection. IndexedDB remembers only the recent directory handle.

          
There is no analytics upload, backend, background runner, WebSocket, or secure credential vault.