Verdict: Nous Research shipped Hermes Agent v0.20.0 ("The Herald Release") on August 3, 2026 — the largest single release in the project's history (more than 3,650 commits across roughly 1,400 merged PRs, closing 1,200+ issues with 650+ community contributors [1]). What changed is not just polish: four new spines were added — conversational voice with barge-in and on-device wake words, an A2A v1.0 plugin that lets Hermes talk agent-to-agent with other frameworks, signed outbound webhooks for event-driven integration, and a grounded-citations skill that prevents fabricated research [1][2]. For builders using open-source agents for daily work, this is the first version of the open-source agent that speaks, cites its sources, and orchestrates other agents without a glue script.
TL;DR — Last verified: 2026-08-04
- Voice is now conversational: speak, get a streaming reply, interrupt mid-sentence by talking back
- A2A v1.0 bundled plugin closes issue #514; works with LangChain, CrewAI, Google ADK and OpenClaw
- Grounded citations skill matches every quote against the real page, plus a fact-checking mode
- Signed outbound webhooks push HMAC-signed events to any HTTP endpoint, no polling
- Desktop app became a platform: sandboxable artifacts with live preview, a real plugin SDK, quick-entry
- CLI wave:
!shell mode,/init,/diff,/context,/focus- License: MIT, free to run on your own hardware;
hermes updateto installPricing/features change often — last checked 2026-08-04.
What is "The Herald Release"?
The Herald Release is the codename for Hermes Agent v0.20.0 (tag v2026.8.3), published on August 3, 2026 by Nous Research, the open-source research lab behind the Hermes models and Hermes Agent runtime [1][2]. Releases are themed and numbered; v0.20 follows a cadence that included v0.17 ("Surface"), v0.18 ("Judgment") and v0.19 ("Quicksilver") earlier in 2026 [1]. The "herald" name is drawn from the Greek god Hermes acting as messenger of the gods — and the release makes him one literally: he speaks (voice), carries word to other agents (A2A), announces events to your systems (webhooks), and cites his sources (grounded research) [1].
The scale alone is worth noting: 3,650 commits, 1,400 merged PRs, 5,200 files changed, 559,000 insertions and 405,000 deletions across 650+ community contributors since v0.19.0 on July 20 [1]. The repository now sits at over 224,000 GitHub stars, making it among the most-starred open-source agent projects on the platform.
What does each headline feature actually deliver?
Below is each pillar of the release, what it does, and why it matters if you build or run AI agents.
1. Conversational voice with barge-in and on-device wake words
Hermes voice mode used to be "speak, wait, listen to one long audio file." v0.20 turns voice into a live conversation. The agent streams the reply as it thinks, you can interrupt mid-sentence by speaking back ("barge-in"), and a wake-word detector runs on-device so you can say "Hey Hermes" from across the room without audio leaving your machine [1][2]. Voice is available on the CLI, the desktop app, and every audio-capable gateway platform — including messaging apps like WhatsApp, Facu, DingTalk, Lark and Q that can automatically reply with text-to-speech when you send a voice note [1][3].
For the technical layer: speech-to-text is fully configurable under hermes tools, OpenAI's gpt-transcribe is supported, and the wake-word engine runs locally for privacy [2]. The use-case that the release notes highlight is hands-free control — "kick off builds and check agents from across the room, no keyboard, no screen" [4]. If you have been using agents tied to a terminal, this is the change that finally lets you drive them like a colleague on a call rather than a slow typist.
This is, in our reading, the keystone capability. Other agents offer voice pipelines, but matching per-word streaming with barge-in and an open-source on-device trigger is rare. We've covered the related computer-control thread before — see our walkthrough on setting up a background AI agent that works while you keep typing.
2. A2A v1.0 — agent-to-agent protocol support
The second pillar ships a bundled plugin implementing the Agent-to-Agent (A2A) protocol v1.0 — effectively closing issue #514, one of the oldest open feature requests in the repository, originally opened in March 2026 [5][1]. The plugin is stdlib-only (no a2a-sdk dependency), lets Hermes discover, talk to, and be driven by other A2A-compatible agents, and works with any A2A-compliant peer — another Hermes instance, LangChain, CrewAI, Google ADK, or OpenClaw [5][6].
The A2A standard itself is an Apache 2.0 protocol donated by Google to the Linux Foundation, designed to complement MCP: where MCP answers "what tools can I use?", A2A answers "who can help me?" Agents discover each other via Agent Cards served at /.well-known/agent.json, negotiate capabilities, and collaborate over standard HTTP/JSON-RPC 2.0, with optional SSE streaming and a structured task lifecycle (submitted → working → completed/failed) [5]. Practically, the bundled plugin exposes Hermes as an A2A server and lets it act as a client to other A2A agents, so multi-agent orchestration across stacks becomes a wire protocol, not a one-off glue script [6].
For builders familiar with the MCP-as-registry pattern, this is the natural pair to it. We have a deeper look at the MCP async tasks story at MCP Tasks: why your AI agents don't support async tools (and how V2 fixes it).
3. Signed outbound webhooks for event-driven integrations
Until now, integrating Hermes into external infrastructure meant polling it or listening on a platform. v0.20 makes Hermes the pusher: it fires signed lifecycle events — session activity, turn completions, tool events — to any HTTP endpoint you register, with [HMAC signatures so your receiver can verify authenticity] [1]. The release examples name CI, home automation, dashboards and any HTTP-speaking service [1][4]. A common pattern this unlocks: when an agent finishes work at 3am you no longer need a poller to find out — your dashboard just gets the event.
This is small but powerful. For builders running a remote agent OS that orchestrates daily work, webhooks turn agent outcomes into OSPF-lite for your own stack — push instead of poll.
4. Grounded citations skill (and fact-checking mode)
Research is where most agents hallucinate. v0.20 ships a new grounded-citations skill that forces every claim to be backed by a verifiable source. The skill matches quotes against the actual text of the cited page (so they cannot be hallucinated), links each citation to the exact evidence, and adds a fact-checking mode that takes any document or claim you hand it and tells you what verifies, what doesn't, and what cannot be checked [1][2].
This is the same idea behind tools likeplexy/perplexity but applied inside the agent, with no reliance on the underlying model's ability to self-flag its own invention. The release authors' framing: "If you use Hermes for research, this is the difference between 'sounds right' and 'provably sourced'" [1]. For small businesses and content teams that rely on agents for research output, this is the closest thing to a non-negotiable capability — a claim with no traceable source is now a tool failure mode, not a model gamble. If you want the larger pattern on agent memory and provenance, our post on the three layers every AI agent OS needs in 2026 goes deeper.
5. The desktop app became a platform (artifacts, plugin SDK, quick entry)
The v0.20 desktop release is arguably a release in miniature. The headline is artifacts: versioned cards rendered in a sandboxed live-preview viewer beside the chat, so generated HTML or apps run safely next to the conversation instead of needing a separate browser tab and root of trust [1][4]. A real plugin SDK also landed, with Kanban as the founding plugin, ctx.download for handing users files, floating pane placement, multiple GUI windows, and a widget-grid layout engine [1][7].
A global-hotkey quick-entry window captures a thought into any session from anywhere in the OS; SSH remote-backend connection mode lets the desktop drive a Hermes instance running on another machine [1][4]. There is also a second 60fps performance wave: streaming cost is now independent of transcript length, dragging works at 60fps with five streaming tabs, and idle CPU is near zero [1]. The result is that the desktop stopped being a chat client and started being a workbench — useful both to power users running multiply-sourced workspaces, and to anyone testing locally generated apps without the "file:// to internet" trust gap that sandboxing eliminates.
6. CLI power-user wave and smarter compression
The CLI got the most concrete set of new commands in the release: ! shell mode (drop in and out of bash), /init (scaffold a skill from a guide), /diff (see what changed), /context (inspect the live context window), and /focus (narrow the model's attention) [1]. There is also mid-turn redirection: if the agent is going wrong, you can type a correction while it works and the active turn reroutes — no need to /stop and re-explain from scratch [4].
Context compression got smarter and gentler: recent conversation always survives, it is progress-aware, and compressed context is presented so most of it shows up in the right form afterward [1]. Smart approvals judge flagged commands themselves by default, so things like deleting a file prompt for confirmation before they execute [1].
7. Speed
The v0.19 Quicksilver release cut first-token time by roughly 80% across every platform — cold-start went from 4.3 seconds down to about 0.9 seconds — and v0.20 continues that trajectory, especially on the desktop app [1][4]. Combined with newer fast models, workflows that took minutes with older models now complete in seconds (the release calls out DeepSeek V4 Flash from late July as the kind of model that amplifies the effect) [1][4]. If you want to see how to actually wire DeepSeek V4 Flash into a free coding setup, we have a step-by-step guide here.
How does v0.20 compare to prior named releases?
v0.20 continues a 2026 pattern of themed Hermes releases, roughly one per month. Understanding the trajectory helps you predict where the next major version may land.
| Tag | Codename | Date | Theme | Most cited additions |
|---|---|---|---|---|
| v0.15.x | "Velocity"/"Patch" | 2026-05 | foundational | Platform stabilization, release cadence begins [1] |
| v0.16.x | "Surface" | 2026-06-05 | UX surfaces | First desktop app drop [1] |
| v0.17.x | — | 2026-06-19 | desktop depth | Live subagent watch-windows, OS notifications, [VS Code themes in-app] [1] |
| v0.18.x | "Judgment" | 2026-07-01 | quality/cleanup | Closes every P0/P1 issue; durable response ledger [1] |
| v0.19.0 | "Quicksilver" | 2026-07-20 | speed | First-token ~80% faster; smart approvals by default; Bitwarden/1Password support [1] |
| v0.19.1 | infrastructure patch | 2026-07-30 | stability catch-up | ~1,000+ PRs rolled into a stable tag for downstream consumers [1] |
| v0.20.0 | "Herald" | 2026-08-03 | voice + interop + proof | Voice, A2A v1.0, webhooks, grounded citations, desktop platform [1] |
The unifying pattern is that the agent's edges matter more than its core: platforms, protocols, voice, and verifiability keep expanding, while the core stays deliberately narrow [2]. The bet is that an agent that talks like a person, proves its work, announces its milestones to external systems, and inter-operates with other frameworks has a longer half-life than one that only optimizes the inner loop.
How to install or upgrade to Hermes Agent v0.20
The release is MIT-licensed and free to run on your own hardware. Existing users update with one command; new users install via the shell installer [1][3].
- Update an existing install: run
hermes updatefrom any terminal where Hermes is on PATH. The agent auto-restarts the gateway if needed [1]. - Fresh install:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash[1]. - Confirm the build:
hermes --versionshould report v0.20.0 (2026.8.3) [1][4]. The release tag was signed with Teknium's verified SSH signature and is published on the GitHub releases page. - Enable voice: open
hermes tools, configure STT and TTS providers (free local options like Edge TTS and local Whisper work; paid providers like OpenAI are optional), and start voice mode. The wake-word detector runs on-device, so audio for that step never leaves your machine [1][2]. - Enable A2A: in the plugin directory,
hermes plugins install <your-a2a-plugin>(or use the bundled implementation) andhermes a2a serve --port 4097. The plugin publishes an Agent Card at/.well-known/agent.jsonand listens for task delegations from any A2A-compliant agent [6][5]. - Turn on webhooks: register your receiver endpoint in
~/.hermes/config.yamlunder thewebhookssection. Verify the HMAC signature on each event using the shared secret.
Note: model usage cost is whatever your LLM providers charge — same as before. Free-tier routing (free local models, Edge TTS) requires zero paid API keys for voice and STT [4][2].
What this means for you
- If you run a small business and want hands-free AI: try voice mode with the on-device wake word. You can finally drive work from across the room, including running builds and asking the agent about state.
- If you operate a fleet of agents (different frameworks): A2A v1.0 lets you connect Hermes to LangChain/CrewAI/Google ADK/OpenClaw peers over a standard protocol, so multi-agent orchestration stops being custom glue.
- If you produce research or client reports: turn on the
grounded-citationsskill and the fact-checking mode. Every claim will link to a verifiable source, and the fact-checker can be pointed at any document your team produces to surface unsupported statements before they reach a client. - If you already have a remote agent OS: outbound webhooks turn agent milestones into first-class events for your CI, dashboards, and home automation — push instead of poll.
- If you live in the desktop app: the artifacts viewer means generated HTML and apps no longer need a separate browser to test, and the plugin SDK opens the door to building your own tool windows beside the chat.
FAQ
Q: Is Hermes Agent v0.20 free to use?
A: Yes. Hermes Agent is MIT-licensed open source, and v0.20.0 is available immediately via the shell installer or pip install hermes-agent. Hosted tiers exist for managed infrastructure, but the core agent is free to run on your own hardware [1][3].
Q: What exactly is A2A v1.0? A: A2A (Agent-to-Agent) is an open protocol (Apache 2.0, under the Linux Foundation) for inter-agent communication. The v0.20.0 plugin lets Hermes discover, communicate with, and be driven by other A2A-compatible agents. It closes issue #514, one of the repository's oldest open feature requests, and works with peers like LangChain, CrewAI, Google ADK, and OpenClaw [1][5][6].
Q: Does voice mode require a paid API key? A: No. The wake-word detector runs on-device, so audio for the trigger step never leaves your machine. STT and TTS providers are configurable — free local options (Edge TTS, local Whisper) exist alongside paid providers like OpenAI's gpt-transcribe [1][2].
Q: How does the grounded-citations skill actually work? A: It matches every quote in generated research against the actual text of the cited page, links citations to the exact evidence, and offers a fact-checking mode that audits any document or claim you provide — reporting what verifies, what fails, and what cannot be checked [1][2].
Q: How big was the v0.20 release window? A: Since v0.19.0 on July 20, 2026: about 3,650 commits, 1,400 merged PRs, 5,200 files changed, 559,000 insertions, 405,000 deletions, 1,200+ issues closed, and 650+ community contributors, according to the official release notes [1].
Q: What happened to the v0.19.1 patch release? A: v0.19.1 (July 30, 2026) was an infrastructure patch tag that rolled up roughly 1,000+ PRs into a stable point for downstream consumers (Docker images, hosted deployments, fresh installs). Its content is fully documented within the v0.20.0 release notes [1][2].
Q: Can I drive Hermes on one machine from another? A: Yes. The v0.20 desktop gained an SSH remote-backend connection mode, letting the desktop UI drive a Hermes instance running on another machine — useful if you keep your agent OS on a server and need a local client for it [1][4].
Discussion
0 comments