Nous Research ships Hermes Agent with real-time voice and agent-to-agent links
The v0.20.0 release adds wake words, grounded citations, signed webhooks, desktop plugins and longer autonomous runs.
By Ryan Merket · Published
Why it matters
Hermes Agent is becoming Nous Research's commercial bridge from open-source model research to subscription software, with voice and interoperability aimed at making agents persistent operating tools.
Nous Research, the open-source AI lab co-founded by Ryan "Teknium" Teknium (@Teknium), shipped Hermes Agent v0.20.0 on August 3rd, adding streaming conversational voice, on-device wake words and a standard protocol for communicating with other AI agents.
https://x.com/NousResearch/status/2084325600643445095
Nous Research announced the release in a two-post thread on X, calling it "The Herald Release." The full GitHub release was signed by Teknium, Nous Research's co-founder and head of post-training, who has steered the Hermes model family from its earlier language-model releases into a persistent agent product.
Teknium says he started Nous Research as an AI research community on Discord in 2022, before Nous Research formalized in 2023. His background includes LLM data engineering at Stability AI and work on character-focused language models at Digi AI. That community-led origin remains visible in Hermes Agent: the code is MIT-licensed, accepts outside contributions and lets users choose models and infrastructure instead of requiring a single provider.
Voice moves to the center
Hermes Agent previously handled spoken interactions as a sequence: record a prompt, wait for a complete response and then play the generated audio. Version 0.20.0 streams speech clause by clause while the answer is generated. Users can interrupt by speaking, causing Hermes Agent to stop playback, listen and redirect the active turn.
Wake-word detection runs on the user's device while waiting for a chosen phrase, according to the release notes. Nous Research says this prevents idle audio from leaving the machine. Different wake words can route requests to separate profiles, while a spoken "stop" command ends a voice session without keyboard input.
The same voice system extends to messaging services. Hermes Agent can transcribe and answer voice notes sent through WhatsApp, Feishu, DingTalk, LINE, QQ, Photon and Weixin. Nous Research also added configurable speech-to-text settings and platform-specific audio formats for replies.
The voice work turns the Hermes Agent product into a more persistent interface across desktops, terminals and messaging services. Hermes Agent already offered memory, scheduling, subagents, browser tools and sandboxed execution. Hands-free control makes those functions available without keeping a terminal window in front of the user.
Hermes Agent gets a standard language for other agents
Version 0.20.0 includes a plugin implementing A2A v1.0, allowing compatible agents to discover Hermes Agent, exchange requests and assign work through a shared protocol. The addition targets developers assembling multi-agent systems from components built on different software stacks.
Nous Research also added signed outbound webhooks. Hermes Agent can send session activity, completed turns and tool events to registered HTTP endpoints, with HMAC signatures that allow receiving systems to verify each message. That gives developers a direct route into CI systems, dashboards and other automation without repeatedly polling Hermes Agent.
Research output received its own verification layer. A new grounded-citations skill ties claims to source material, checks quotations against page text and links citations to supporting evidence. A related fact-checking mode classifies claims based on what the agent could verify. These are Nous Research's product claims, and the release notes do not provide independent accuracy measurements for the citation system.
Longer runs and a broader desktop surface
Nous Research raised Hermes Agent's default tool-iteration ceiling from 90 to 500, a material change for autonomous tasks that require repeated searches, terminal commands or file edits. Tool failures now return recovery information: truncated terminal output can be written to a file, patches can identify whitespace mismatches, searches can probe for near matches and file writes can verify the saved contents.
Users can also correct Hermes Agent during an active turn without discarding completed work. New CLI commands run shell operations, generate an AGENTS.md file from a project scan, show code changes and break down context-window usage. An import command can migrate configurations from Claude Code or Codex CLI.
The desktop application now supports versioned artifacts with sandboxed live previews, multiple windows and a plugin SDK. Kanban is the first bundled plugin. A global hotkey opens a quick-entry window for sending a note into an existing session from elsewhere in the operating system.
Nous Research's official materials use different snapshots to describe the release's size. The GitHub notes list about 3,650 commits, 1,400 merged pull requests, 1,200 closed issues and more than 650 contributors since v0.19.0. The graphic attached to the X announcement lists 3,427 commits, 1,300 pull requests and 1,168 issues. Both sets of figures should be treated as approximate indicators of the development window.
Open source feeds a paid product
Hermes Agent is free under an MIT license, while Nous Research sells subscriptions through Nous Portal, which bundles model access and tools such as web search, image generation, text-to-speech and a hosted browser. Nous Research offers free and paid Portal tiers, creating a commercial layer around the open-source agent.
That business model is drawing investor attention. TechCrunch reported on July 13th that Nous Research was finalizing a round of at least $75 million led by Robot Ventures, with Union Square Ventures participating, at a $1.5 billion valuation. The report cited people familiar with the transaction, and Nous Research did not confirm the round.
Version 0.20.0 shows where Nous Research is directing that product push: persistent voice access, interoperable agents, longer autonomous runs and integrations that place Hermes Agent inside existing software operations. The open repository supplies distribution and contributors. Portal gives Nous Research a way to charge users who want the models and infrastructure packaged together.