AgentDeck - Mission Control for AI Coding Agents

Developers run multiple AI agents but manage them blind. AgentDeck brings visual status, one-tap controls, and haptic alerts to your CLI coding agents.

Comment

Inspiration

I run multiple Claude Code sessions daily - one refactoring auth, one writing tests, one debugging deployments. But managing them? Pure chaos.

I'm constantly alt-tabbing between terminals, scrolling up to find missed approval prompts, losing track of which agent needs me. The AI is smart. My workflow isn't.

When I saw Logitech's DevStudio challenge, I realized: the MX Creative Console is the perfect mission control for headless AI agents.

What it does

AgentDeck transforms your MX Creative Console into a command center for AI coding agents like Claude Code.

LCD Keypad - Live Dashboard

  • Top row: Each button represents an agent session with live status
    • 🟢 Green = Idle (ready for input)
    • 🟡 Yellow = Working (thinking/coding)
    • 🔴 Red = Waiting (needs your approval)
  • Middle row: Context-aware actions (Approve, Reject, Pause)
  • Bottom row: Workspace commands (Undo, Show Diff, New Task)

Dial - Contextual Control

  • When agent is idle → scroll command history
  • When agent is working → scrub through plan steps
  • When reviewing → navigate diffs file-by-file

Haptics (MX Master 4)

  • 3 pulses = agent needs approval
  • 1 long pulse = task completed
  • Rapid vibration = error occurred

No more tab hunting. No more missed prompts. You stay in flow.

How I built it

AgentDeck consists of two components:

  1. Bridge Service: A daemon that monitors terminal sessions, parses Claude Code output, and tracks agent state (status, current task, files modified, approval requests)

  2. Logi Actions Plugin: Built with the Actions SDK in C#, receives state from the bridge and:

    • Updates LCD button displays with agent status
    • Sends commands back to terminals on button press
    • Triggers haptic feedback on MX Master 4 for notifications

The bridge communicates with the plugin via local WebSocket, enabling real-time updates without polling.

Challenges

  • Parsing unstructured terminal output: Claude Code doesn't have a formal API, so detecting state changes requires pattern matching on CLI output
  • Multi-session coordination: Tracking which terminal maps to which agent slot, handling session starts/stops gracefully
  • LCD refresh timing: Balancing update frequency with visual stability

What I learned

  • The Actions SDK is surprisingly powerful - dynamic folders, multistate actions, and haptics open up interaction patterns I hadn't considered
  • CLI-based AI agents are the future, but their UX is stuck in the past
  • Physical controls create a fundamentally different relationship with software than keyboard shortcuts

What's next

  • Support for additional agents (Codex CLI, Gemini, Aider, Goose, etc.)
  • Token usage and cost tracking display
  • Git-aware undo integration
  • Actions Ring shortcuts for quick commands
  • VS Code extension for tighter terminal integration

Built With

Share this project:

Updates