Hermes Agent v0.20.0 Herald Release: Voice, A2A & More

Hermes Agent v0.20.0, codenamed the Herald Release, is Nous Research's 3 August 2026 update to its open-source, MIT-licensed AI agent framework. By the project's own numbers, it is the biggest single release since Hermes Agent launched: roughly 3,650 commits, around 1,400 merged pull requests, about 5,200 files changed, some 559,000 lines added and 405,000 removed, and roughly 1,200 issues closed, all contributed by more than 650 individuals since v0.19 Quicksilver shipped on 20 July.

Where Quicksilver was almost entirely a plumbing release (cold-start speed, approval safety, secrets management), Herald is a genuine capability release. Four features anchor it: real-time conversational voice with barge-in and on-device wake words; Agent-to-Agent (A2A) v1.0, a standard protocol for talking to other agents regardless of what they are built on; signed outbound webhooks that push events instead of waiting to be polled; and a grounded-citations research skill that backs claims with verifiable, quote-matched sources. Alongside those four, the desktop app gained artifacts, a plugin SDK and a large performance pass, and the CLI picked up half a dozen new power-user commands.

• Best for: developers and power users who already run Hermes across a terminal, desktop app and messaging platforms and want it to speak, coordinate with other agents, and push events rather than being polled.
• Headline additions: streaming voice with barge-in, A2A v1.0, signed webhooks, grounded citations, desktop artifacts and a plugin SDK.
• Scale: ~3,650 commits, ~1,400 merged PRs, ~5,200 files changed, 650+ contributors since Quicksilver.
• Main caveat: most of these are net-new surfaces (voice, A2A, plugins) rather than refinements, so expect the usual early-release rough edges even in a mature project.

Hermes Agent is built by Nous Research, an AI research organisation better known previously for its Hermes line of fine-tuned open-weight language models. Hermes Agent is a separate, newer product: an agent harness, the orchestration layer that gives a language model memory, tools, scheduling and a place to run, rather than a model in its own right. Point Hermes at GPT-5.6, Claude Opus 5, Gemini 3.6 Flash or a local open-weight model, and the harness behaves identically; the reasoning quality underneath is a function of whichever model you chose.

The project is free and open-source under the MIT licence and runs as a CLI, a terminal UI, a native desktop app, or a server-hosted gateway that bridges into more than a dozen messaging platforms, sharing one persistent identity and memory across all of them. That "one agent, many surfaces" design is Hermes' clearest differentiator from more IDE- or terminal-centric competitors like Claude Code. For the fuller introduction to the product and its architecture, see our original Hermes Agent explainer, and for the immediately preceding release, our Quicksilver review.

Nous Research published the Herald changelog on GitHub on 3 August 2026, tagged v2026.8.3. The scale dwarfs even Quicksilver's already-large release: roughly 5,200 files changed against Quicksilver's 2,465, and around 559,000 lines added against Quicksilver's 300,000. Where Quicksilver's workstreams were mostly about making the existing product faster and safer, Herald's are almost entirely new surface area: voice infrastructure, an entirely new inter-agent protocol, a webhook delivery system, a citation-verification skill, and a desktop app rebuilt around a plugin architecture.

The name fits: a herald announces and coordinates, and this release is fundamentally about Hermes talking, being heard, and talking to other agents, rather than only executing tasks silently in a terminal.

The headline feature is real-time conversational voice. Hermes now streams text-to-speech clause by clause as the underlying model generates its reply, rather than waiting for the full response before speaking, which cuts the perceived latency between asking a question and hearing an answer begin. More notably, it supports barge-in: a user can interrupt Hermes mid-sentence just by speaking. The agent stops, listens, and the model is explicitly told the user cut in, so the next turn accounts for the interruption rather than restarting from scratch.

On-device wake words let a user define their own open-vocabulary trigger phrase, "Hey Hermes" or anything else, with detection running locally so no audio leaves the machine while the agent is simply listening for its name. Saying "stop" ends the voice session hands-free on every supported surface. Voice works across the CLI, the desktop app, and gateway adapters, with platform-aware delivery that optimises the audio format for WhatsApp, Feishu, DingTalk, LINE, QQ and Weixin specifically, rather than shipping one generic format everywhere. The underlying speech-to-text infrastructure was also unified and now supports configurable providers, including OpenAI's transcription API, instead of a single hard-coded backend.

This closes a gap that had become conspicuous: most of Hermes' direct competitors, from OpenAI's ChatGPT voice mode to Google's Gemini Live, had shipped real-time voice well before this, and Hermes' text-only interaction model was one of the few places the open-source project visibly lagged the frontier consumer products, despite matching or beating them on agentic capability.

The second pillar is Agent-to-Agent (A2A) v1.0, a standard interoperability protocol that lets a Hermes agent discover and communicate with other A2A-compatible agents, regardless of what those other agents are built on. Nous Research is explicit about the distinction from Model Context Protocol (MCP): MCP connects an agent to tools; A2A connects agents to other agents, letting heterogeneous multi-agent systems coordinate over a shared wire protocol while each side keeps its own independent stack, memory and model choice.

Notably, this closes GitHub issue #514, reportedly the single oldest open feature request in the Hermes Agent repository, a detail that signals how long multi-agent coordination has been on the community's wishlist relative to when it actually shipped. It also lands the same month Anthropic's Claude Code added its own cross-session messaging tools (ListAgents and SendMessage), suggesting agent-to-agent communication is becoming a convergent feature across the entire harness category rather than one project's idiosyncratic bet: see our companion piece on Claude Code's cross-session messaging for how Anthropic approached the same underlying problem.

Signed outbound webhooks

Hermes can now push signed lifecycle events, session activity, completed turns, and tool events, to registered HTTP endpoints, using HMAC signatures so a receiving system can verify a payload genuinely came from that Hermes instance. This flips the integration pattern from repeatedly polling Hermes for status to receiving events the moment they happen, which matters for anyone wiring Hermes into CI/CD pipelines, home-automation systems or live dashboards, use cases where polling either wastes resources or introduces unacceptable lag.

Grounded citations

The new grounded-citations skill backs every factual claim in a research answer with a verifiable source, using quote matching against the actual source text rather than a plausible-sounding but unverified reference. A companion fact-checking mode can audit an existing document or claim and report whether each assertion checks out, fails verification, or is simply unverifiable from available sources. This targets one of the most persistent trust problems with agentic research tools: a fluent answer that cites nothing, or cites something that does not actually say what the agent claims it says.

The desktop app received the release's largest single workstream. It now supports artifacts: versioned cards with a sandboxed live preview rendered in a side viewer, similar in spirit to Claude's own artifacts feature but running inside Hermes' own desktop shell. A genuine plugin SDK ships alongside it, with a Kanban board as the founding plugin, plus a ctx.download API, floating panes, custom widgets and support for multiple simultaneous GUI windows, turning the desktop client from a single chat surface into an extensible application host.

Smaller but practical additions include a global-hotkey quick-entry window for capturing a thought without switching focus, an SSH remote-backend connection mode so the desktop app can drive a Hermes instance running on a remote server, and composer upgrades: a file/folder/link attachment picker, an undo stack, and iMessage-style emoji reactions on messages. Performance work delivered a genuine 60fps while streaming across five simultaneous tabs, a workload that reportedly used to visibly stutter.

On the command line, the release adds several power-user commands: !command for instant shell execution without leaving the agent session, /init to scan a project and generate or update an AGENTS.md file, /diff to show staged or session changes, /context for a breakdown of what is consuming the context window, /focus for a reduced-output view, and Ctrl+S to stash a half-written prompt. A new hermes import-agent command migrates an existing Claude Code or Codex CLI setup into Hermes in one step, explicitly lowering the switching cost for users coming from a competing harness.

Beyond new features, Herald continues the performance discipline Quicksilver established. Cold-start time falls again, from roughly 14 seconds in earlier configurations down to about 1.8 seconds; config reads are reported as 54 times faster after eliminating unnecessary deep-copying; and update operations run two to six seconds faster for no-op cases. Streaming cost is now independent of transcript length, addressing a scaling problem where longer conversations previously made every subsequent streamed token more expensive to render. Idle CPU usage in the background is reported as near zero.

On the reliability side, the tool-calling iteration limit was raised from 90 to 500, letting an agent work through substantially longer multi-step tasks before hitting an artificial ceiling. Context compression received a further overhaul on top of Quicksilver's work: proactive tool-result pruning, per-turn micro-compaction, and a guarantee that the most recent user messages are always preserved regardless of how aggressively older context gets compressed. Tool self-recovery also improved: truncated terminal output now spills to a readable file instead of silently cutting off, patch detection identifies edits that were already applied so the agent does not reapply them, and failed searches probe for near-matches rather than simply reporting nothing found.

As with any vendor-reported figure, these numbers come from Nous Research's own release notes rather than independent benchmarking, and "cold start" measures time-to-first-response rather than end-to-end task completion, which still depends primarily on the model connected. The consistency of reported gains across CLI, desktop, gateway and cron surfaces is nonetheless a reasonable signal that the underlying fixes are structural rather than cherry-picked.

The scale of this release did not happen in a vacuum. In July 2026, Nous Research finalised a $75 million Series B at a $1.5 billion valuation, led by Robot Ventures with Union Square Ventures participating, roughly fifteen months after Paradigm led the company's $50 million Series A. The round is reported to support expansion of both Hermes, the agent platform this article covers, and Psyche, Nous Research's separate decentralised AI infrastructure network. A jump from a small open-source research outfit to a $1.5 billion valuation in just over a year is a useful piece of context for why the Herald Release ships four genuinely new subsystems (voice, A2A, webhooks, citations) at once rather than the more incremental pacing of earlier releases: a much larger, better-funded contributor and engineering base is now behind the project.

A release this large inevitably reads better on paper than it behaves in the first few weeks of real use, and the honest caveat is that three of the four headline features, voice, A2A and the plugin SDK, are genuinely new surfaces rather than refinements of something already battle-tested. Voice interfaces in particular are notoriously hard to get right: barge-in detection has to distinguish a genuine interruption from background noise, wake-word detection has to avoid false triggers, and clause-by-clause TTS has to sound natural rather than choppy. Nous Research's own release notes do not claim these are fully mature; expect the kind of early rough edges that are normal for a first major version of a voice stack, even inside an otherwise mature project.

A2A v1.0 carries a different kind of real-world caveat: its usefulness depends entirely on how many other agents and platforms adopt the same protocol. A single agent that can speak A2A is not yet useful if nothing else in a user's stack speaks it back; the feature's value will grow (or not) based on ecosystem adoption over the following months, not on anything Nous Research alone controls. As with Quicksilver, none of Herald's plumbing and feature work changes the underlying reasoning quality of whichever model is connected: pairing Hermes with a weaker model will still produce a noticeably less reliable agent, voice, A2A and webhooks notwithstanding.

Hermes Agent the software remains free and open-source under the MIT licence, installable via a shell installer or pip install hermes-agent, with no subscription or per-seat fee for the framework itself. As before, the two real cost centres are hosting (Nous Research continues to cite a roughly $5-a-month VPS as workable, or a near-free serverless backend such as Daytona or Modal while idle) and the underlying LLM API usage, billed by whichever provider is connected. Herald adds new default model options to choose from, including Gemini 3.1 Pro, the newly discounted Gemini 3.6 Flash, Claude Opus 5, and DeepSeek V4 Flash 0731, alongside a new Buzz gateway (Block's Nostr-based messenger) and a modernised Vercel AI Gateway integration.

As with Quicksilver, Nous Research has not published one fully itemised public breakdown of every optional Nous Portal tier's exact price and credit allowance; independent estimates of realistic all-in running cost (VPS plus model usage, with or without a portal plan) still land somewhere in the $5-to-$80-a-month range depending on model choice and usage volume. Treat that range as a third-party estimate rather than an official published figure, and check the current Nous Portal plan list directly before budgeting.

• Voice is a first major version: barge-in, wake-word accuracy and natural-sounding streamed TTS are all genuinely hard problems; expect rough edges relative to years-mature consumer voice products.
• A2A's value depends on ecosystem adoption: a single agent speaking a standard protocol is only as useful as how many other agents and platforms adopt the same standard.
• Harness, not intelligence: none of Herald's new surfaces change the reasoning quality of whichever underlying model you connect; results remain a function of model choice.
• Cost is still not fully published: the software is free, but hosting plus model API usage, and an optional portal subscription without complete published tier pricing, make total monthly cost harder to predict up-front than a flat-fee competitor.
• Self-hosting overhead persists: choosing a terminal backend, a model provider, and now potentially plugin and A2A configuration, is meaningfully more setup than opening a polished single-vendor consumer app.

Herald narrows a real gap Hermes had against single-vendor products: voice. OpenAI's ChatGPT and Google's Gemini had offered real-time voice interaction for years before this release, while Hermes remained text-first. Combined with A2A, Hermes is now making an explicit multi-agent-ecosystem bet that neither Claude Code nor OpenAI's Codex app has matched with a public, standard protocol at the time of writing, though Claude Code's own new cross-session messaging tools address a related, narrower version of the same coordination problem within Anthropic's own ecosystem rather than across vendors.

The core trade-off from earlier reviews still holds: Hermes asks more of the user upfront (choosing a model, a terminal backend, and now potentially plugins and voice configuration) in exchange for model-agnosticism, no vendor lock-in, and now genuine cross-agent interoperability. A developer who has already picked one frontier model and wants the smoothest single-vendor experience will likely still find a purpose-built tool from that model's own maker more polished out of the box. A developer who wants one agent identity that talks, coordinates with other agents, and follows them across a terminal, a desktop app and messaging platforms now has a meaningfully stronger case for Hermes than before Herald shipped.

Use the Herald Release if you already run Hermes and want voice interaction, cross-agent coordination via A2A, or event-driven integrations via webhooks, or if you are evaluating open-source, model-agnostic agent harnesses and value an active project shipping this much surface area this quickly.

Wait a release or two if you specifically need production-grade voice or A2A today: both are first major versions inside this release, and the more conservative move for anyone running Hermes in a setting where reliability matters more than novelty is to let the community shake out the inevitable early bugs before depending on either feature.

The Herald Release is a different kind of update than Quicksilver. Where Quicksilver made an already-good harness faster and safer, Herald makes a bet on where agent products are heading next: talking instead of only typing, coordinating with other agents instead of working in isolation, and pushing events instead of waiting to be polled. Backed by a fresh $1.5 billion valuation and the largest contributor push in the project's history, Nous Research is clearly not treating Hermes as a side project.

For anyone already invested in the open-source, model-agnostic agent approach, Herald is worth pulling to get voice and A2A into the toolkit early, with the usual caveat that first versions of genuinely new subsystems deserve a slightly longer trust-building period than a routine point release. For anyone comparing Hermes against Claude Code or Codex, this release closes Hermes' most visible remaining gap against consumer-facing competitors and adds a capability, standard cross-agent communication, that neither has matched yet in the same form.