What happened. An open source plugin for Claude Code plays an audio clip from Mr. Meeseeks (from the Rick and Morty cartoon) when the agent finishes a task and awaits instructions, or when it asks for approval. The GitHub repository collected 130 points on Hacker News and 54 comments within hours. The plugin filters notification events by type, stays silent during autonomous work (auto-accept, subagent), and only sounds when you’re actually needed. It supports macOS without additional installations, Linux with ffmpeg or mpg123, Windows via PowerShell.

Why it matters to you. When an agent works for thirty minutes in the background and you switch tabs, missing the moment it stalls on an approval means grinding everything to a halt. This plugin signals a real need: agentic tools need UX designed for long sessions, where the user isn’t watching the screen continuously. It’s not a critical feature, but it recognizes that an agent isn’t a command that finishes in two seconds: it’s a collaborator that can work for minutes or hours, and the interface must adapt to that rhythm.

If you want to try it. The plugin installs directly from Claude Code with /plugin marketplace add thephw/claude-meseeks, then /plugin install mr-meeseeks@claude-meseeks. It works immediately on macOS. On Linux you need an audio player (ffplay or mpg123). Precompiled binaries are included—no compilation needed.

In detail

How it works

The plugin hooks into Claude Code’s notification events through hooks/hooks.json, registering on Notification and UserPromptSubmit. When the agent finishes a task and waits (notification_type: idle_prompt), a clip plays from the audio/done/ folder (“All done!”, “Ooh yeah!”). When it asks for approval (confirmation notification_type), it pulls a clip from audio/asking/ (“Can you help me?”). Autonomous work—auto-accept, background agent, auth refresh—stays silent.

Clips are embedded directly in the Go binary (meeseeks), distributed precompiled for macOS, Linux, and Windows in the bin/ folder. If the right binary for your platform isn’t there, the launcher tries to compile from source; if Go is also missing, it stays silent without blocking Claude Code. Playback runs in detached mode and never halts the prompt. On macOS it uses afplay (built-in), on Linux it looks for ffplay or mpg123, on Windows it uses PowerShell Media.SoundPlayer.

Why it’s interesting

It’s not a critical tool, but it signals a need many users have already hit and that enterprise agentic tools will need to solve: managing attention across long sessions. An agent that works for 20 minutes then stalls on an approval creates a worse experience if you have to keep the tab open to catch it. As we covered on July 12, Anthropic’s analysis of 1.2 million sessions confirms agents mostly serve administrative and writing work: repetitive tasks requiring multiple steps that you want to fully delegate. To truly delegate, though, the tool needs to alert you when it needs you, without you checking constantly.

The project is completely open source (license not specified in the README, worth verifying), well documented, and shows how to extend Claude Code through the plugin system without modifying the app. The 54 Hacker News comments mostly discuss practical utility and suggest variants (native notifications instead of audio, volume options, custom clips). The reaction is positive: many users have the same problem, and the plugin solves it without complexity.

Limitations and context

It’s a workaround for an interface not designed for long sessions. The ideal solution would be a native, configurable system notification with states visible even when the window is minimized. The plugin requires an audio player (built-in on macOS, needs installing on Linux), and on Windows the PowerShell solution isn’t always reliable. There’s no way yet to customize clips without recompiling the binary.

But the point isn’t technical perfection: it’s that a small tool, published by an independent developer, attracted immediate attention because it solves a real need that official tools haven’t addressed yet. It’s a signal for those building agentic tools: UX doesn’t end when the agent starts—it begins there.