Safety
Know when an agent needs stopping.
Session keys show how full the conversation is and how fast it burns tokens, warn you when a session looks off, and let you stop the agent's current turn from the deck. All of it is on by default and can be turned off in Setup.
Stop a turn: three taps.
Tap a session key three times within two seconds to stop the agent's current turn. The session stays open, so you can tell the agent what to do instead. If VibeDeck is holding a waiting approval for that session, it is declined first.
Stop works in iTerm2, Ghostty, tmux and herdr panes, Windows Terminal (x64) and Codex Direct Control. Other hosts, status-only agents and the Claude VS Code panel show that Stop isn't available instead of sending keystrokes.
Context at a glance.
The session key fills up as the conversation grows. It turns amber from 80% and shows a red percentage from 95%, so you can wrap up or start a fresh session before the agent runs out of room. Claude Code, Codex and Pi, when the agent reports its context.
Cache hit rate.
Next to the context percentage, the key shows ↻ and the session's prompt-cache hit rate: how much of the conversation the AI provider reused from its cache instead of processing it again. A high rate usually means faster replies and lower token costs. A sudden drop shows the cache was lost, for example after a long break or a model switch.
The rate covers the whole session, so one slow request after a pause doesn't hide a healthy cache. Want the key to fill by cache hit rate instead of context? Turn on Fill by cache hit rate in Setup → Safety; both numbers stay on the key. Claude Code (terminal and Claude Desktop), Codex and Pi, when the agent reports its token usage. New in 1.0.6.0.
Token burn pulse.
A working Claude Code or Codex key pulses faster when the agent burns tokens faster than your usual pace. Slow means it is sipping; fast means it is burning through them.
Early warnings.
A red border on the key tells you when:
- Too many subagents: Claude Code reaches your limit of subagents running at once (10 by default).
- Runaway burn: a Claude Code or Codex session burns far above its own pace (3× by default).
- Busy without progress: a Claude Code or Codex session keeps spending tokens with no tool activity for 30 minutes (by default). Long reasoning can trigger it too, so check before you stop anything.
Tap the key to dismiss a warning. Warnings never stop anything for you.
Subagents on the key.
Claude Code keys show how many subagents are running, and their tokens count toward the session.
Your rules.
Open Setup to turn Stop, context fill, the burn pulse or any warning on or off, and to set your own limits: subagents (2–50), runaway pace (2–10× your own) and minutes without progress (10–120). New Stopped and Alert sounds tell you when a turn stopped or a warning appeared.
Which agents.
- Stop (three taps): input-capable sessions in iTerm2, Ghostty, tmux, herdr, Windows Terminal (x64), and managed Codex.
- Context fill: Claude Code, Codex, Pi.
- Cache hit rate: Claude Code, Codex, Pi.
- Token burn pulse: Claude Code, Codex.
- Runaway and no-progress warnings: Claude Code, Codex.
- Subagent count and warning: Claude Code.
Status-only agents (Gemini CLI, Kilo, Qwen Code, Hermes, Kimi Code) show status only. See compatibility for every agent and host.
Private by design.
Token counts come from the agents' local session files on your computer, never your conversations, and they stay on your computer. See the privacy policy.
Try it on your own agents.
Safety is included with VibeDeck 1.0.5.0. Install it, connect Claude Code or Codex, and watch the keys. The trial needs no card and ends without a charge.