Federico Ulfo
The problem with running three coding agents at once isn't the agents — it's that each one silently stops and waits for you, and you don't find out until you tab back to that terminal. Clippy attacks exactly that gap, and nothing else.
A macOS desktop app that watches Claude Code and Codex sessions and gives each one a small animated buddy that perches on the corner of the terminal or editor window that owns it. The buddy is quiet while the agent works, shows a one-line activity status, and surfaces a card the moment a human decision is needed.
This is the part worth understanding, because it explains both the strengths and the limits:
localhost:43117. The installer registers curl hooks in ~/.claude/settings.json and ~/.codex/hooks.json, and the agent POSTs its lifecycle events to it.PermissionRequest becomes an Allow/Deny card, AskUserQuestion becomes clickable buttons, ExitPlanMode becomes a plan review, Stop becomes a review-the-work card.
Founder, Engineer
New York City
Everything is localhost. No cloud relay, no telemetry, and — importantly — nothing is ever auto-approved.
The failure mode is the old behaviour. Because the hooks time out into the normal terminal prompt, the worst case of Clippy being broken, closed, or asleep is that you're back to where you started. That is the correct design for a tool that sits in the approval path, and it's the reason it's safe to leave installed.
Per-session identity does real cognitive work. Each session gets its own colour, character and pet name derived from the session ID. After a day of it, "the orange one is stuck" is a genuinely faster thought than reading four terminal titles.
It's honest about scope. It doesn't try to be a dashboard, an orchestrator, or a second agent. It's a notification surface with buttons.
@anthropic-ai/claude-agent-sdk plus an OAuth token or ANTHROPIC_API_KEY — more setup than the rest of the app, and worth treating as a separate feature.If you run more than one coding agent on a Mac, install it — it's free, MIT-licensed, and the downside risk is bounded by design. If you're on Linux or Windows, or you run agents mostly in CI rather than in front of you, there's nothing here for you yet. The interesting question it raises is whether "human attention router" becomes a standard layer in agent tooling; Clippy is the smallest possible version of that idea, which is why it works.
Disclosure: Clippy is built by AI Socratic. We use it daily, which is why it's reviewed here — read the verdict with that in mind.
Sources: