hermes v0.20 has landed. Codenamed Herald, it is the release where Hermes finds its voice: real-time conversation you can interrupt mid-sentence, a custom wake word that works from across the room, and spoken replies inside WhatsApp, DingTalk and LINE. Stack on grounded citations, an agent-to-agent plugin, sandboxed Desktop previews and a first token that now arrives in 0.9 seconds instead of 4.3, and you have one of the densest changelogs Hermes has ever shipped. This page is the complete run-through.
Quick refresher if you are new here: Hermes is the open-source AI agent that lives on your computer and talks to any model you point it at. V0.20 follows the Quicksilver release — the V0.19 update that was all about speed and fleet management — and this one arrived with 647 contributors behind it.
Julian Goldie tested Herald on the day it dropped, putting it through the same Goldie Bench hands-on testing he runs on every major agent and model release. Everything below comes from that testing and the release notes — and updating takes about a minute through Agent OS, which we will cover further down.
What Shipped in hermes v0.20 (Herald)
Herald is a dense release, so here is the full what-changed list in one table. Every item gets more detail below.
| Feature | What it does |
|---|---|
| Talk to Hermes | Real-time voice conversation. The voice speaks while the response is still generating, and you can interrupt it mid-sentence just by talking. |
| Custom wake word | Say your chosen phrase from across the room to activate Hermes hands-free. |
| Voice on every platform | Send a voice note on WhatsApp, DingTalk, LINE and other chat platforms; Hermes replies with automatic, platform-aware text-to-speech. |
| Grounded citation skill | Research with sources cited on every response, deep-research style, to avoid hallucinations. |
| Hermes Desktop artifacts | View everything Hermes creates across platforms in one place, with sandboxed live previews for HTML apps. |
| A2A plugin | Hermes can discover, talk to, and be driven by other A2A-compatible agents. |
| Mid-turn correction | Type a correction while Hermes works and the active turn redirects — no /stop and re-explain. |
| Compression improvements | Progress-aware context compression; your recent conversation always survives. |
| Smarter approvals | Risky actions, like deleting a file, get checked with you first. |
| Outbound webhooks | Ping your other systems the moment a piece of work finishes. |
| Speed work | First token in 0.9 seconds instead of 4.3; the telemetry gate is 54x faster. |
| Security and fixes | Security updates and bug fixes across the agent. |
The Voice Upgrade
The headline act. Three voice features shipped together in Herald, and combined they change how the agent feels day to day.
Talk to Hermes: Real-Time Conversation
This is proper real-time voice, not the old speak-wait-listen loop. The voice starts speaking while the response is still generating, so there is no dead air while a wall of text renders somewhere. Better still, if Hermes heads off in the wrong direction, you interrupt it mid-sentence simply by speaking — exactly like cutting in on a phone call. It stops, listens and adjusts.
A Custom Wake Word, From Across the Room
You can now set your own wake word and trigger Hermes hands-free from the other side of the room. Herald bundles the recent wake word update into the main release, so the whole flow is one polished loop: say the word, speak the task, get on with your day.
Voice on Every Platform
Voice is no longer a desktop-only trick. Send Hermes a voice note on WhatsApp, DingTalk, LINE or another supported chat platform and it replies with automatic text-to-speech, tuned to whichever platform you are on. If Hermes already runs inside your messaging apps, it now feels properly native.
📺 Watch: Hermes Agent V0.20 Just Changed AI Agents Forever!
Grounded Citations: Research Without the Guesswork
The new grounded citation skill gives you research with sources cited on every response, in the same spirit as deep research tools. Ask a factual question and you get the answer plus where it came from. If you use agent output in client work or published content, this is the practical way to avoid hallucinations sneaking through.
Hermes Desktop: Artifacts and Sandboxed Previews
Hermes Desktop gains an artifacts view: one place to see everything Hermes has created for you, across every platform you use it on. The bigger deal is sandboxed live previews — HTML apps that Hermes builds now run inside a sandbox before they touch anything real, which is a genuine safety upgrade rather than a convenience. Julian is honest that Hermes Desktop did not look great in earlier versions; in Herald it is much faster and considerably better looking.
📺 Watch: Run Hermes Agent Free Forever, Here's how...!
The A2A Plugin: Agents That Talk to Agents
The new A2A (agent-to-agent) plugin lets Hermes discover other A2A-compatible agents, talk to them, and even be driven by them. In plain terms, Hermes can become the front door to your whole agent stack — or a worker inside someone else's. Julian's verdict after testing: this is the single thing to watch in the entire release. If the A2A protocol keeps improving at this rate, controlling all your agents from one place gets far easier.
The Quality-of-Life Wins
Beyond the headliners, Herald ships smaller changes you will feel every session:
- Mid-turn correction. Type a correction while Hermes is working and the active turn redirects on the fly. No more typing /stop and re-explaining the whole task from scratch.
- Compression improvements. Context compression is now progress-aware, and your recent conversation always survives. Long sessions stop losing the thread.
- Smarter approvals. Risky actions — deleting a file, for instance — now get checked with you before they happen, instead of after.
- Outbound webhooks. Hermes can ping your other systems the moment work finishes, which is the missing link for chaining it into bigger automations.
- Security updates and bug fixes. The usual under-the-hood pass, and worth having on an agent with this much access to your machine.
📺 Watch: Higgsfield MCP Makes Hermes Agent Unstoppable
The Speed Numbers
Herald is faster in ways you can actually measure. The first token now arrives in 0.9 seconds instead of 4.3, so the wait before a response starts has essentially collapsed. The telemetry gate — one of the internal bottlenecks — is 54x faster. And Hermes Desktop is much quicker than it was in earlier versions.
Where it really sings is alongside the new generation of fast models. DeepSeek V4 Flash shipped just days before Herald, and in Julian's hands-on testing the pairing was brilliant: running /learn to teach Hermes a guide and turn it into a skill file came back in seconds. A fast agent on a fast model is a different experience from either one alone.
How to Update to Hermes V0.20
The update is painless, and there are two routes:
- Through Agent OS: open your Agent OS dashboard, head to the Manage section, and click Update.
- Through the terminal: run hermes update and let it do its thing.
That is genuinely it. Herald installs over your existing setup, so your skills, memory and platform connections carry across. New to the stack? The Agent OS guide linked earlier walks you through setup from zero.
If you want Hermes V0.20 wired into a money-making system, check out the AI Profit Boardroom — the full Agent OS setup plus daily step-by-step tutorials are inside. → Put the Herald release to work in your business
Hermes V0.20 FAQ
What is the Hermes V0.20 Herald release?
Herald is version 0.20 of Hermes, the open-source AI agent that runs on your computer and talks to any model you point it at. Headline features: real-time voice conversation, a custom wake word, voice replies on chat platforms, grounded citations, an A2A agent-to-agent plugin and major speed gains. 647 contributors worked on the release.
How do I update to Hermes V0.20?
Open the Manage section of Agent OS and click Update, or run hermes update from the terminal. Your existing skills, memory and connections carry over automatically.
What is the difference between Quicksilver and Herald?
Quicksilver (V0.19) was the speed and fleet-management release. Herald (V0.20) builds on it with the voice suite, grounded citations, the A2A plugin, Desktop artifacts and further speed work — first token in 0.9 seconds versus the old 4.3.
Can Hermes reply with voice on WhatsApp?
Yes. Send a voice note on WhatsApp, DingTalk, LINE or another supported platform and Hermes replies with platform-aware text-to-speech. No setup beyond updating to V0.20.
What is the A2A plugin in Hermes V0.20?
It is agent-to-agent protocol support. Hermes can discover other A2A-compatible agents, communicate with them, and be driven by them — the groundwork for controlling a whole fleet of agents from one place. Julian rates it the most important long-term piece of the release.
The Bottom Line
Herald is the release where Hermes stops being something you type at and becomes something you talk to. Real-time voice with interruptions, a wake word from across the room, voice notes on the platforms you already live in — plus grounded citations, sandboxed previews for anything it builds, and a 0.9-second first token that makes the whole agent feel instant.
The sleeper, though, is A2A. Voice is the feature you will show people; agent-to-agent control is the one that changes what Hermes becomes over the next few releases. If that protocol keeps improving, running every agent you own through one front door stops being a demo and starts being the default. The update takes a minute through Agent OS — 647 contributors did their part, so go collect yours.
