For a while the interesting OpenClaw work on this house lived only in a private Unraid repo: a Discord bot that sits in a voice channel all day, a local Whisper gate so ambient chat never hits a cloud STT meter, YouTube audio that ducks when the bot talks, a JSON-RPC bridge to Red-DiscordBot, and a thin tool that reads ThermoPro room sensors over BLE.
That is now a public repo — pjmagee/openclaw-plugins — because the plugins were already running in production, and the useful part is not a hello-world SDK sample. It is how they fit together on a box that is also the NAS, the game host, and the thing that should not stream a public voice channel to the cloud because someone forgot a setting.
This post is about that system, not a rewrite of each README.
Why extract them
OpenClaw plugins load inside the Gateway process. Ours started as files under ~/.openclaw/extensions/ on Tower, next to operator docs, 1Password readers, and an hourly patch guard. That is a terrible place to keep software you want other people to install: one git pull of the fleet repo is not a release, and a homelab checkout is full of things that must never ship.
So we cut five packages out, each with a manifest, a configSchema, and a tarball on GitHub Releases (thermopro-v0.1.1, yt-media-v0.1.0, and so on). Config stays in openclaw.json / env. The repo has no secrets, by design. Install is pinned to a release URL so an unrelated plugin bump cannot change the bytes you just downloaded.
They still run on the same Unraid host first. Public is a snapshot of what we operate, not a parallel codebase we hope is similar.
The five pieces
| Plugin | What Chillbot actually uses it for |
|---|---|
| thermopro | thermopro_read — office/bedroom temp, humidity, battery from TP3x BLE beacons |
| redbot-control | redbot_* tools, only when someone names Red. Ordinary “play this” must not go there |
| yt-media | Default music: real YouTube audio on the same Discord AudioPlayer as Grok TTS |
| whisper-wake-gate | The spec + tests for local name-gating. Not installed as its own plugin on Tower |
| xai-realtime-voice | The live voice stack: Grok Voice, local wake gate (same bytes as the module above), mic presence, working cue, music ducking |
whisper-wake-gate is the odd one. On the fleet it is not enabled as a standalone plugin. xai-realtime-voice vendors a copy of wake-gate.ts and CI fails if the two files diverge. The standalone package exists so the matching rules, fail-closed policy, and debug tools (wake_gate_status / wake_gate_check) have a home with a test suite. The voice plugin is what actually hears Discord.
How a day on Tower looks
Chillbot is an always-on Discord account. Primary brain is Grok. Speech is xAI realtime voice (sirius). It sits muted in a voice channel and listens locally.
Ambient audio is buffered, cut into utterances, and sent to a local OpenAI-compatible Whisper endpoint. We run two on purpose:
- npu-stt on the Intel Arrow Lake NPU — cheap enough to listen all day (~6.7× lower power than CPU for this job).
- speaches on the 3090 — used for bulk transcription via a skill, not for the always-on gate.
If the transcript does not contain a wake name (and the alias table does not save a Whisper mis-hear like “killbot” for “chillbot”), the audio is dropped. Nothing goes to xAI. If we marked the gate required and that local backend is down, the bot goes deaf. Fail closed. A silent stream of a public VC to a cloud STT is not an acceptable fallback.
Say the name and the mic opens, the utterance is handed to Grok Voice (local transcript armed as a timer fallback so a missing cloud STT event cannot leave the bot mute-and-confused), a quiet “working” loop can play if the reply is slow, then the gate re-arms. Conversation mode (name gate off) exists for a session and does not persist across rejoin or restart. Forgotten open-mics should not last days.
Music is the same connection. media_play resolves YouTube with yt-dlp, ffmpeg emits Ogg/Opus (Discord-ready without a Node opus encoder), and the track shares one AudioPlayer with TTS. When the bot speaks, the track pauses and seeks back after about 1.2s of silence. Cookies for Premium/age-gated media come from 1Password through a refresh command — the jar is never in config or logs. After a successful play the agent is told not to speak a confirmation: that sentence would immediately duck the track it just started.
While a track is playing, yt-media sets OPENCLAW_MUSIC_PLAYING. The wake gate then requires the name again. Ambient “yeah play the next one” in a busy VC must not skip your queue. Those env vars (OPENCLAW_MUSIC_PLAYING, OPENCLAW_VOICE_SELF_*) are the cross-plugin bus. Plugins load as separate module trees; process.env is the contract that survives a reload.
Red is a second bot on the same Discord. Its official JSON-RPC has no auth and only listens on loopback inside Red’s container, so a socat sidecar republishes it on 172.17.0.1:6134 — the Docker bridge gateway, which does not change when containers recreate. The plugin’s usageHint is load-bearing: the model only touches redbot_* when a human says Red. Snowflakes stay strings end-to-end so JavaScript cannot round an 18-digit guild id.
Sensors are the simple end of the same pattern. thermopro-ble runs on the host network (BLE sockets do not see hci* from a normal container netns), publishes JSON on :8102, and the plugin is an HTTP client. “How warm is the office?” is a tool call, not a scrape.
Almost every companion address we use is 172.17.0.1, not a container name or a DHCP’d container IP. The gateway container sees the host there. Recreates do not silently break the tools.
Things we would do again
No secrets in the plugin repo. Config schema and env names, yes. Tokens, cookie jars, processing-loop audio we did not author — no. Tower still has a private cue file; the public package ships a recipe script instead.
Fail closed on the wake gate. Required + backend down = drop mic. We would rather a deaf bot than a leak.
Compare mode before cutover. When the Python gate was replaced with speaches, both ran on live audio with the old verdict authoritative and divergence logged. The counters are still in the module.
Honest about patches. Parts of the wake-gated Discord flow live in @openclaw/discord’s compiled dist. The voice package ships the same idempotent applier we run in production, plus a table of what each patch does. Without them you still get an xAI realtime provider; you lose name-gating, capture-during-playback, and the local audio hook. An hourly guard on Tower re-applies after image pulls. That is not elegant. It is documented.
One AudioPlayer, tagged resources. TTS, the working cue, and YouTube all share a player. Every stop() checks identity. Nothing force-stops audio it did not start. That rule is why ducking and “working” do not murder a track.
What we did not publish
The fleet still has operator docs, 1Password readers, Discord account config, and the processing-loop opus. Those belong next to the machine. The plugins belong in a repo you can read in one sitting — which is the point of the SDK examples, too. openclaw.plugin.json, registerTool, text results, no mystery.
If you want the install commands and schemas, start at the repo README. If you want to know whether they work: they already do, on the same host that is serving this thought from a different pipeline.