Hermes Agent 0.20.0 adds streaming voice, wake words, and grounded citations
TL;DR
On August 3, 2026 Nous Research released Hermes Agent v0.20.0, the Herald release, tagged v2026.8.3. Voice replies stream and can be interrupted, wake words run on device, and a grounded-citations skill checks quotes against the page they came from.
- Status
- Live
- Released
- Plans
- Hermes Agent
- Source
- —
What changed
The release named Hermes Agent v0.20.0 (v2026.8.3) is dated August 3, 2026. GitHub published that tag on August 3, 2026. The notes call it the Herald release and count the window since v0.19.0 at about 3,650 commits, about 1,400 merged pull requests, and 650+ contributors.
Highlights in the notes:
- Streaming voice. Hermes speaks clause by clause as the reply streams. You can interrupt by talking. It stops, listens, and the model is told you cut in. Silence detection is supposed to keep it from talking over you. This works in CLI voice mode, on the desktop, and through gateway adapters.
- Wake words. An open-vocabulary wake phrase starts listening. Detection runs on the device, so audio does not leave the machine while it waits. Different wake words can route to different profiles. Saying "stop" ends the voice chat on every surface.
- Voice notes on messaging platforms. A voice note on WhatsApp, Feishu, DingTalk, LINE, QQ, Photon, or Weixin is transcribed and answered. Speech-to-text is its own tools category, and the notes name OpenAI gpt-transcribe as a supported engine. A spoken-text preprocessor strips markdown, code, and URLs before text-to-speech.
- Grounded citations. The
grounded-citationsskill makes research where claims are backed by a source. Quotes are matched against the page text. A fact-checking mode does the same for a document or a claim you hand it, and reports what checked out, what did not, and what could not be verified. - Outbound webhooks. Hermes can push signed lifecycle events, including session activity and turn completions, to your systems. The notes contrast that with polling or listening on a platform.
Why it matters
Voice in 0.20.0 is a conversation you can interrupt, not a file you wait for. Wake-word detection stays on the device until the phrase hits, which is the privacy line in the notes. "Stop" is the hang-up, including when your hands are not on the keyboard.
Grounded citations are the research control. The skill is supposed to match a quote to the page instead of accepting a plausible sentence. Use it when a memo has to show where a line came from. The notes do not say it is on for every answer. It is a skill you use.
Outbound webhooks are how another system hears that a turn finished without scraping the gateway. The notes say the events are signed.
The install tag is v2026.8.3. The version name is v0.20.0.
Frequently asked questions
Does the wake word send audio to a server while it is waiting?
The v0.20.0 notes say detection runs on the device, so no audio leaves the machine while it waits.
How do I stop a voice chat?
The notes say saying "stop" ends the voice chat on every surface.
Are citations automatic on every reply?
The notes describe a grounded-citations skill that checks quotes against page text and can fact-check a document you provide. They present it as that skill, not as a new default for all answers.