Hermes Agent v0.20.0: streaming voice with barge-in, outbound webhooks, desktop plugin SDK
Large release centred on real-time conversational voice: Hermes now speaks clause-by-clause as it generates, can be interrupted mid-sentence, and responds to on-device wake words across CLI, desktop, and messaging gateways. It also gains signed outbound webhooks for pushing lifecycle events to external HTTP endpoints, a grounded-citations research skill, and a desktop plugin SDK with sandboxed artifact previews.
PUBLISHED2026-08-03
OBSERVED2026-08-11
AGE8d
SOURCES1
- Streaming TTS speaks clause-by-clause instead of one file after full generation; barge-in stops playback when you talk and tells the model it was cut off; busy-aware silence detection
- Open-vocabulary wake phrases detected on-device (no audio leaves the machine while waiting); multi-profile wake-word routing; spoken "stop" ends voice chat on any surface
- Voice notes transcribed and answered on WhatsApp, Feishu, DingTalk, LINE, QQ, Photon, Weixin; platform-aware auto-TTS delivery (opus where required, captions attached)
- STT made configurable: own
hermes toolscategory, GUI toggles, dashboard dropdowns, unified language resolution, OpenAIgpt-transcribesupport; one shared spoken-text preprocessor strips markdown/code/URLs across all TTS providers - New
grounded-citationsskill: quotes matched against actual page text, citations link to exact evidence, plus a fact-checking mode reporting verified / contradicted / unverifiable - Signed outbound webhooks push session activity, turn completions, and tool events to registered HTTP endpoints with HMAC signatures — no polling required
- Desktop: versioned artifact cards with sandboxed live preview in a right rail, plugin SDK (Kanban as first plugin),
ctx.download, floating panes, multiple windows, global-hotkey quick-entry - Release intro also lists A2A v1.0 agent-to-agent messaging, CLI commands
!shell mode,/init,/diff,/context,/focus, and smarter compression - Rolls up the v0.19.1 infrastructure patch tag; ~3,650 commits, ~1,400 PRs, ~1,200 issues closed since v0.19.0
COMMUNITY
No curated reactions recorded for this event. Facts and takes are kept in separate layers — community context is added by hand, never blended into the record above.