AgentDeck - Command Your AI Fleet

Mission control for your AI coding fleet. Glanceable status tiles, one-tap approvals, dial-driven diff scrubbing, and haptic alerts - across every VS Code, Windsurf, and Cursor window.

Comment

Inspiration

We run multiple AI coding agents in parallel every day. One refactors auth. One writes tests. One chases a deployment bug. The agents are brilliant. Managing them is chaos.

We tab-switch to check "is it done yet?". We scroll back to find a missed prompt. We lose track of who's waiting on us. We miss BELs. We alt-tab to the wrong window. We lose flow.

The agents are from 2026. The interaction model is from 1975.

And we're not alone. Every developer we talk to is wrestling with the same thing. The AI era flipped our workflow overnight. We've gone from writing every line ourselves to orchestrating fleets of agents. The tooling hasn't caught up. We still babysit each agent in its own terminal, burning cognitive fuel on window management instead of shipping. Agents got 10x more capable. Our interface to them got 10x more expensive to operate.

We wanted to build the thing we needed - and that every AI-native developer needs. A physical control surface that turns multi-agent chaos into something you can glance at, feel, and steer with one hand. Something that gives developers their flow state back. Being a great engineer now means being a great conductor.

When we saw the Logitech DevStudio challenge, it clicked. The MX Creative Console is the perfect mission control for headless AI agents. A glance tells us who needs us. A tap tells them what to do. A buzz on the MX Master 4 tells us when to look up.

The console should be the remote control. The editor is the screen.



What it does

AgentDeck turns your MX Creative Console + MX Master 4 into a physical command center for AI coding agents. It works with Claude Code, Gemini CLI, Codex, Aider, and OpenCode inside VS Code, Windsurf, or Cursor.

AgentDeck on Logi Options+



Three layers of physical interaction

Layer 1 - LCD Dashboard: Glanceable Status (The dynamic folder - AgentDeck)

A Dynamic Folder takes over all 9 LCD buttons. Up to 6 agent tiles per page, page through the rest with the dial. Manage an unlimited fleet, not a fixed slot count. Each tile shows live status: 🟢 running, 🟡 pulsing needs input, 🔴 error, ⚪ idle/ready.

Main dynamic folder with agents listMain dynamic folder with agents list and different status



  • Single tap -> raises the right editor window and focuses that agent's terminal.
  • Double tap -> opens a per-agent Skills page (End, Commit, Review, Clear, Diff, Mode, Continue).

Single agent skills page



Tap the Sessions tile for fleet-wide actions across every window:

Sessions screen



Layer 2 - Assignable Dialpad & Buttons: Fleet Controls Every dial, button, and Actions Ring slot is an assignable action in Logi Options+. Users wire up their own loadout. Our daily driver:

  • Top-left dial -> Pause All (Ctrl+C every running agent)
  • Top dial -> Continue All (approve or nudge every agent at once)
  • Roller -> Permission Mode (cycle ask / auto / plan via Shift+Tab)
  • Main dial -> Contextual + Agent Dial (rotate to cycle agents, press to focus; in Diff mode, rotate scrubs changed files, press toggles file <-> hunk)
  • Bottom-left button -> Show Actions Ring
  • Bottom-right button -> Next Waiting (jumps to the next agent needing input)

Standalone actions in Logi Options+



To have your own action tiles, drag any AgentDeck action onto any surface - keypad, dialpad, or MX Master 4 Actions Ring.

Standalone actions in Logi Options+


Layer 3 - MX Master 4: Haptic Feedback The mouse buzzes on status changes. sharp_collision when an agent needs input. completed when it finishes. angry_alert on error. Ambient awareness - no need to look at the console.



New Agent flow with git worktree isolation

Tap NEW on the LCD Dashboard -> pick from 6 agent types. Tweak settings on the same screen via CONFIGS: WORKTREE, Permission Mode, Effort level.

When worktree is ON (🟢green), the extension runs git worktree add and spawns the agent in an isolated branch and directory (e.g., agentdeck/claude-myproject-a1b2). Agents on the same repo no longer clobber each other. Worktrees share the .git object database, so creation is instant and lightweight.

New agents screenNew agents config



Skills page: prompt-as-action (double tap any active agent tile)

Double-tap any agent tile to open its Skills page. Each tile sends a real message to that agent's terminal:

  • Commit - asks the agent to commit its work
  • Clear - kills and relaunches the agent with a fresh worktree
  • Review - Review code changes from current branch
  • Diff - opens the agent's changed files in VS Code's diff viewer
  • Continue - smart: approves when waiting, sends "continue" when idle
  • Mode - cycles permission mode via Shift+Tab (ask / auto / plan)
  • End - kills the agent session

Note: After a double-tap, skill tiles are locked for ~700ms to prevent an accidental third tap from firing a skill.

Single agent skills page



Standalone actions on Logi+

The Logi+ plugin ships many custom actions you can attach to separate tiles:

Standalone actions on Logi+



Quick Prompt - your own skill slots: Bind any button to a Quick Prompt with your own text in Logi Options+. Tap to send it to the active agent. Build your own muscle-memory skills - "run tests", "explain this file", "write a commit message" - and fire them with one press.

Quick Prompt action on Logi+



Launch Agent - create new agent within your IDE: Bind any button to a Launch Agent with options of agent type and worktree config. Tap to create a new agent with that predefined config.

New agent action on Logi+



Effort + Permission Mode dials. Two purpose-built dial actions:

  • Effort Level - rotate to cycle /effort low / medium / high on the selected agent. Press to reset.
  • Permission Mode - rotate to walk through ask / auto / plan via Shift+Tab.

Dial configs



Multi-window fleet (one of the most important features)

Every VS Code / Windsurf / Cursor window runs its own extension on its own WebSocket port (:9999-:10008). The Logi plugin discovers all of them and merges the agent lists into one dashboard. Commands route to the right window via a port prefix on each agent's ID. Launch a NEW agent -> it spawns in the last-focused window. Tap on an agent tile, it brings up the correct window with correct terminal, ready for review and input

Zero-config auto-attach (so you don't even need to launch the agent from the console)

Type claude, gemini, codex, aider, or opencode in any terminal. The extension detects it via VS Code's Shell Integration API, auto-attaches, and streams output for full status detection. For older terminals without shell integration, a backup output-signature matcher spots agent branding ("Claude Code", "Gemini CLI", "Aider v...") and promotes them too. No launcher, no wrapper, no config.

Optional local AI status classifier

When the heuristic parser is uncertain, an optional local Ollama model (defaults to qwen2.5:0.5b, ~1GB RAM) breaks the tie. Off by default in IDE settings. Zero cloud, no API costs. Async and debounced with silent fallback, so it never slows detection down.

IDE settings



How We built it

AgentDeck is two components, zero daemons:

  1. VS Code Extension (TypeScript) - the brain. Spawns agents via node-pty + vscode.Pseudoterminal, detects status, serves a WebSocket server, handles diffs, worktrees, and window focus.
  2. Logi Plugin (C#, Actions SDK) - the face. Renders LCD tiles, handles input, manages multi-window connections, drives haptics.

Killing the bridge

v1.0 had a separate bridge daemon. v2.0 doesn't. The extension is the server. No tmux, no external binary, no launch agent. When VS Code closes, the agents close - a reasonable trade because the console user always has an editor open.

4-layer status detection

Detecting agent state from TUI output is the single hardest problem in this project. We ended up with a layered pipeline, each layer faster and more specific than the next:

Layer Mechanism Latency
1. Escape sequences BEL (\x07), OSC 9/777 Instant
2. Heuristic patterns Stripped ANSI + per-agent regex 2s poll
3. Silence detection Time since last PTY chunk 2-10s
4. AI classifier Optional local Ollama fallback Async

BEL is the killer feature. Claude Code emits it the moment it needs approval, so the tile turns 🟡 within a single data chunk - no polling, no pattern matching. The TTY spec has carried this signal since 1975; we just finally gave it a UI.

Silence detection solved the "is it done?" problem that regex couldn't. TUI agents stream spinner frames every ~100ms while working. The absence of data is the signal. Per-agent thresholds (2s for Claude, 10s for Aider) map cleanly onto each agent's streaming behavior.

Pseudoterminal, not proposed APIs

Our first attempt used VS Code's proposed terminalDataWriteEvent API to intercept output from native terminals. It worked perfectly in Extension Development Host. It was silently blocked in production - VS Code scans for enabledApiProposals at load time and refuses to activate. Even stripping the field but leaving the .d.ts in the VSIX tripped the check.

The fix: node-pty + vscode.Pseudoterminal (stable API). We spawn a real PTY (xterm-256color, proper cols/rows) and bridge it into VS Code's terminal UI. Output flows to both the visible terminal and our status parser. TUI agents get real isatty(), we get unfiltered data. And node-pty is loaded from VS Code's bundled node_modules.asar - no native binary in the VSIX.

Multi-window routing

The port-scanning handshake took a few iterations. Each extension instance probes :9999 -> :10008 sequentially and claims the first free port. Agent IDs get prefixed with the port (w9999-agent-1) so they stay globally unique across windows. The plugin's BridgeMultiClient opens all 10 sockets in parallel, merges state, and routes commands by parsing the prefix.

For window focus (tapping an agent tile should raise the right editor window), osascript was the trap. Windsurf and Cursor register their process name as "Electron" - tell application "Windsurf" doesn't work. The fix: match by bundle identifier (com.exafunction.windsurf, com.microsoft.VSCode, com.todesktop.230313mzl4w4u92) via System Events, then frontmost + AXRaise.

Hardware-free Web Simulator

A browser-based simulator (bun dev → localhost:8888) renders the full LCD, dial, and button interactions. Lets us develop, test, and demo AgentDeck without the physical console plugged in - and made the entire project buildable before hardware was wired up.



Challenges We ran into

Proposed API lockout. We spent a full day convinced we could obfuscate the onDidWriteTerminalData property access. VS Code scans compiled output too. The only real answer was a full architectural pivot to Pseudoterminal.

TUI output is a nightmare. Agents redraw entire screens with ANSI sequences. Regex that worked on Monday was broken by a UI tweak on Thursday. Silence detection became the safety net - if the stream stops, the agent is done, regardless of what's in the buffer.

Logi SDK gotchas. The SDK is powerful but the docs leave edges. A few that cost us hours:

  • PluginDynamicFolder with NavigationArea.None still reserves position 0 for Back - we only had 8 usable buttons, not 9.
  • ButtonActionNamesChanged() only re-renders if the parameter names change, so we had to bake an epoch counter into every parameter to force refresh on view switches.
  • The MX Creative Console requests bitmaps at Width116 (116x116), not the 80x80 most docs imply. Our first tiles rendered upscaled and blurry.
  • MakeProfileAction("tree") dropdowns silently disable the Save button unless you call AddLevel() exactly twice - single-level trees look correct but won't save.

No hold/long-press in the SDK. The MX Creative Console fires RunCommand with no duration metadata, so we rolled our own double-tap detection (~400ms window) + a 700ms view-switch cooldown to prevent a third tap from firing a skill action right after transitioning.

VSIX runtime deps. .vscodeignore excludes node_modules/** by default. The extension activated silently with no error - commands just... didn't exist. The fix was one line: !node_modules/ws/**.



What We learned

  • Physical controls change your relationship with software. A tile that turns yellow and buzzes your mouse is cognitively different from a notification you have to tab over to see. Ambient awareness beats explicit checking every time.
  • The absence of signal is a signal. Silence detection solved more status bugs than a week of regex tuning.
  • Terminal escape sequences are an underused API surface. BEL, OSC 9, OSC 777 - the TTY spec already had the hooks we needed. We didn't need a new protocol; we needed to listen to the one that's been there since 1975.
  • Stable APIs beat clever APIs. Pseudoterminal is less elegant than onDidWriteTerminalData but ships to real users.
  • The MX Creative Console is a real input device, not a toy. The dial for diff scrubbing feels genuinely better than any keyboard shortcut we've used.



What's next

  • Claude Code Hooks integration - replace BEL/heuristic parsing with the structured Hooks API (PreToolUse, Notification, Stop -> direct status). Same approach for Gemini's --output-format stream-json and OpenCode's SDK/SSE.
  • Default .lp5 profile - pre-assign the Agent Deck folder on install so users don't have to wire it up manually.
  • Full Actions Ring loadout - Approve / Reject / NextWaiting / Kill / OpenTerminal as standalone Actions Ring commands on MX Master 4.
  • Cost + token tracking - per-agent spend visible on the dashboard.
  • Session forking - branch an agent's conversation into a new worktree with one tap.
  • Windows & Linux - plugin is cross-platform; needs PowerShell window focus (Windows) and Logi Options+ for Linux.



Notes



Built With

Share this project:

Updates