Hermes Agent Update v0.20 is MASSIVE! (Herald Release)
What the Herald release actually changes
Hermes Agent v0.20, called the Herald release, is a big update on paper. The headline feature is a reworked voice mode with a wake word: say "Hey, Hermes" and it spawns a new session, hands-free. Under that, there's a new grounded citation skill, an agent-to-agent protocol, signed webhooks, desktop artifacts, and a pile of CLI quality-of-life improvements.
Here's our practical call after walking through it: the citation skill and the CLI changes are the real reasons to update. Voice mode is nice to have, but it is not how we would run serious work.
Voice mode: handy, but keep your expectations low
The new voice mode works best in the Hermes desktop app, where all the toggles live. The wake word makes it genuinely hands-off, and the pitch is clear: you're on the couch, in the shower, or away from your desk, and you still want to talk to your agent.
In practice, we found it awkward. While we were recording, the live voice conversation kept picking up everything we said, which is annoying if you're already sitting at your computer. You stop it by saying "end." And even with the improved pipeline, our honest position hasn't changed: typing is still the best way to work with Hermes. It's more accurate, it gives you time to think, and proper spelling and grammar are simply more efficient input for a model that processes tokens. Voice is fine for blurting out small stuff. Do not use it for big tasks, and definitely not for coding.
Setup notes: in settings, enable speech-to-text echo transcripts. OpenAI is the best provider, but the free local option is decent. If you run Hermes on a VPS like we do, voice mode is a pain. A headless CLI environment has no mic or port audio, and the workarounds are fiddly. Stick to the desktop app, and if you want your VPS agent there, connect it over Tailscale and take the session over on desktop. You'll need a mic, obviously.
The grounded citation skill is the sleeper feature
This is the upgrade we actually care about. If you use Hermes for research, you know the failure mode: once the context window fills up halfway, or you're ten turns deep, the agent starts drifting and sometimes hallucinates plausible-sounding claims.
The grounded citation skill fixes that by attaching claims to verifiable sources. Quoted text gets matched against the actual source material, and citations point to supporting evidence. There's also a fact-checking mode: hand it a document or a claim, and it tells you what checks out, what doesn't, and what couldn't be verified. In effect, this makes Hermes behave a lot more like Perplexity, without paying for a separate Perplexity subscription. If your work is research-heavy, this alone justifies the update.
For the tinkerers: A2A, webhooks, and desktop artifacts
With the bundled A2A v1.0 plugin, Hermes can now discover, talk to, and even be driven by other compatible agents. If you're building multi-agent systems across frameworks, that gives you a standard wire protocol. Signed outbound webhooks are the companion piece: Hermes can push signed lifecycle events (session activity, turn completions, tool events) to HTTP endpoints, and HMAC signatures let the receiving system verify they genuinely came from Hermes. That makes CI dashboards, home automation, and monitoring integrations much cleaner.
On the desktop app, generated apps, pages, and documents now show up as versioned artifact cards with sandboxed live previews in a right-rail viewer. There's a plugin SDK with Kanban as the founding plugin, plus a global quick-entry hotkey to capture thoughts into any session. The desktop app is starting to feel less like a chat window and more like an agent workbench.
CLI quality of life
For terminal people, this release is genuinely good. The new ! shell mode runs a shell command instantly without burning a model turn, still behind the normal approval gate. New slash commands include /init for generating agents.md, /diff for reviewing changes, /context for inspecting what's actually filling your context window, and /focus for a reduced-output view.
The standout is mid-turn redirection: if Hermes heads the wrong way, you can type a correction while it works instead of stopping and restarting. Think of it as an abrupt version of /steer — use redirection when it's fully off track, /steer when it's just drifting. There's also a hermes import agent command to migrate a Claude Code or Codex CLI setup into Hermes.
Performance got attention too: the default tool-calling iteration limit went up roughly six times, cold starts are faster, and long sessions should stay coherent longer thanks to redesigned compression with proactive tool-result pruning and per-turn micro-compaction.
Bottom line: update for citations and the CLI. Try voice mode if you're often away from your desk, but keep typing for anything that matters.