Pipecat Peekaboo demo runs a macOS assistant on local STT with cloud LLM/vision workers
Peekaboo is a desktop-operator proof of Pipecat workers plus local hear-side capture — closer to an end-user assistant than changelog multi-worker rows, without claiming a shipped product SKU.
Also UseHearSee
Details2 sources · 2 X posts
Kwindla (@kwindla) showcases Aleix Conchillo Flaqué's (@aconchillo) experimental Peekaboo macOS voice assistant: local transcription and wake-word monitoring, cloud LLM and screen vision by default, with a WebView Pipecat JS client on a native macOS audio transport and workers on a shared bus.
Peekaboo macOS voice assistant demo
.@aconchillo has been hacking on an experimental macOS voice assistant. Audio transcription and wake word monitoring run locally. LLM and screen vision use cloud models by default, but of course you can swap in a local model/endpoint.
"Peekaboo, tell me when the build finishes."
"Peekaboo, somebody invited me to a meetup about inference optimization tonight, but I can't find the message. Was that in Slack, email, WhatsApp, or iMessage?"
"Peekaboo, did we do the evals last week for the new Nemotron model on boule, pancake, flatbread, or the dgx spark? I think we forgot to push to the repo."
https://x.com/kwindla/status/2098880411279802391
https://x.com/kwindla/status/2098880412869357723
Pipecat Context Hub 0.8.0 moves to the MCP 2.x SDK and indexes Pipecat UI instead of Voice UI Kit
Pipecat UI becomes the indexed React source before pipecat 1.11.0 switched its init clients to it (2026-09-17-pipecat-1-11-0-pipecat-ui, W38); the mcp<2 pin in Pipecat's [mcp] extra is a concrete install conflict.
Details1 source
Pipecat Context Hub 0.8.0 migrates the server to the mcp 2.x SDK and swaps pipecat-ai/pipecat-ui in for pipecat-ai/voice-ui-kit among the default sources.
Pipecat Context Hub 0.8.0
SierraPlatformlaunch2026-09-10
Sierra ships multimodal agents mixing voice, text and visuals
Treats visuals as part of the voice-agent turn, not a hand-off to a separate app — the post is feature narrative without named customer go-lives or latency numbers.
Details2 sources · 1 X post
Sierra's multimodal agents keep voice, text and on-screen visuals in one conversation so customers can talk, view options and tap without restarting the session.
Multimodal agents
Hume AIModel providerother2026-09-10
Hume publishes a multi-axis voice cloning leaderboard
Public multi-axis cloning scores rather than one Elo. VoxCPM2 led same-speaker (4.21), Cartesia sonic-3.6-beta led naturalness (4.36), Inworld TTS-2 led quality (4.61).
Details1 source · 1 X post
Hume published a Voice Replication Leaderboard measuring how well 11 TTS models clone 25 reference voices across identity, quality, naturalness, and objective similarity.
Voice Replication Leaderboard
PolyAIPlatformlaunch2026-09-09
PolyAI launches Wren, an agent that improves dialog agents from live conversations
Moves PolyAI from shipping Dialog-RSN-1 models into continuous agent optimisation on live voice and digital traffic, with named customers (Golden Nugget, Hawksmoor, Simplyhealth) already using it.
Details1 source
PolyAI launched Wren inside Dialog Studio: an agent that reviews every production conversation, proposes fixes and improvements, tests them, and shows metric movement for human approval.
Wren
Deepgram Voice Agent can defer function calls until end of turn
Closes a speculative-reply footgun where end_call or booking tools could fire before Flux confirmed the user had finished speaking.
Also Think
Details3 sources
Deepgram added defer_until_eot on Voice Agent functions so irreversible tool calls wait for turn confirmation, and emits FunctionCallCancelled when barge-in or a resumed turn voids a speculative call.
Voice Agent defer_until_eot
Pipecat demo plays chess with Muse Spark over a distributed-agents bus
A concrete multi-machine voice+vision+chess harness shows distributed agents beyond docs examples — PhoneLLM on the voice loop while heavier Muse Spark / Stockfish work stays off the critical path.
Also UseThinkSee
Details3 sources · 2 X posts
Kwindla (@kwindla) demos a weekend chess setup where a Pipecat bot on a DGX Spark (Nemotron ASR, PhoneLLM, Pocket TTS) talks to distributed coding and chess agents over a Pipecat message bus, with Muse Spark deciding moves via Stockfish/Maia and a vision pipeline on an RTX 5090.
Muse Spark chess distributed-agents demo · vendor Meta
Weekend project: play chess with Muse Spark running on a DGX Spark with code Muse Spark wrote.
Setup:
- projector
- HDMI, camera, mic, and display server (Raspberry Pi)
- pipecat bot (DGX Spark)
- vision pipeline (RTX 5090)
- coding agents running on the Pi, the DGX Spark, and the machine with the 5090
- vibe-coded chess backend with Muse Spark making the natural language, contextual decisions and calling into Stockfish, and Maia (DGX Spark)
- everything is duct-taped together with ssh, tmux, and redis
I talk to the pipecat bot, which is:
- NVIDIA Nemotron ASR
- Pipecat PhoneLLM
- Kyutai/Gradium Pocket TTS
All the other parts of the system are distributed agents talking via a pipecat message bus.
I can talk to the bot and it responds directly, or sends stuff to the coding agents or the chess backend. It can send messages directly to components it knows a lot about, or just put generic messages on the bus. The pipecat bot system instruction does know about the chess environment and message structures for moves, etc. (But, it would be much more elegant to load that on demand as a skill.)
Lots of room to make this much faster, cleaner, more extensible, more robust, etc. But it works! You can play full games against Muse Spark. Or ask it for tutorials. Or play against another person and get commentary from any of the models.
https://x.com/kwindla/status/2096840980997771449
https://x.com/kwindla/status/2096840982386057518