Nous Research adds local wake-word activation to Hermes Agent

Nous Research, the open-source AI lab co-founded by Ryan Teknium, added local wake-word activation to Hermes Agent on July 29th, allowing users to open a session and issue a spoken command without touching a keyboard. The default phrase is "Hey Hermes," and the feature works across the command-line interface, terminal UI, and desktop application, but is disabled by default. Hermes Agent, which grew out of internal research tooling, now combines terminal-native interface with persistent memory, reusable skills, and voice activation.

Nous Research https://nousresearch.com , the open-source AI lab co-founded by Ryan Teknium @Teknium1 https://x.com/Teknium1 , added wake-word activation to Hermes Agent https://hermes-agent.nousresearch.com/ on July 29th, letting users open a session and issue a spoken command without touching a keyboard. https://x.com/nousresearch/status/2082509593280688317?s=46 https://x.com/nousresearch/status/2082509593280688317?s=46 The lab announced the feature in a post on X https://x.com/nousresearch/status/2082509593280688317?s=46 . The default phrase is "Hey Hermes," and the feature works across Hermes Agent's command-line interface, terminal UI and desktop application. It is disabled by default. Hermes Agent grew out of the founders' own research workflow. Nous Research co-founder Jeffrey Quesnelle @theemozilla https://x.com/theemozilla told Practical AI https://practicalai.show/357/transcript that Teknium first developed the agent as internal tooling for researchers automating model work. Nous Research later opened the repository as the project expanded beyond its original model-training use cases. That history remains visible in the product. Hermes Agent combines a terminal-native interface with persistent memory, reusable skills, scheduled jobs, browser tools, coding capabilities and messaging integrations. Voice activation gives that developer-oriented system an interface associated with consumer assistants while retaining access to terminals and local models. How the local listener works According to Nous Research's wake-word documentation https://hermes-agent.nousresearch.com/docs/user-guide/features/wake-word , enabling the listener starts a lightweight hotword detector on the user's default microphone. When Hermes recognizes the trigger phrase, the detector pauses, releases the microphone, opens a fresh session and records one spoken command. The command is transcribed, passed to the agent and followed by a spoken response before the detector resumes listening. The bundled setup uses openWakeWord and includes a pre-trained "Hey Hermes" model. Hermes Agent also supports sherpa, which can recognize a user-specified phrase without training, and Picovoice's Porcupine engine. Users can adjust detection sensitivity and require several consecutive matching audio frames to reduce accidental activations in rooms with background speech. The privacy boundary is narrower than fully local voice processing. Wake-word detection stays on the device, and Nous Research says no audio leaves the machine before the trigger is recognized. After activation, the command enters Hermes Agent's normal speech-to-text and model pipeline. Users can keep transcription local with faster-whisper and run a local model, while configurations using hosted speech or AI providers can send command data off-device. Hermes Agent also restricts the feature to local interfaces with an available microphone. It does not apply to Hermes agents operating through Telegram, Discord or other messaging gateways. Only one local Hermes surface can control the microphone at a time. Nous Research is widening Hermes Agent's distribution Voice activation extends a rapid push beyond the command line. The Hermes Agent site https://hermes-agent.nousresearch.com/ offers desktop installers for macOS and Windows, Linux installation, messaging connections and cloud access through Nous Portal. Nous Research sells hosted tiers with credits for models and tools, giving the lab a paid layer around the MIT-licensed agent. The open-source project already has substantial developer distribution. The Hermes Agent repository https://github.com/NousResearch/hermes-agent had more than 222,000 stars and 42,500 forks on GitHub as of July 29th. Those figures measure developer attention rather than active usage or revenue, but they give Nous Research a large funnel for desktop and hosted products. That funnel is drawing investor interest. TechCrunch reported on July 13th https://techcrunch.com/2026/07/13/hermes-agent-maker-nous-research-in-talks-for-new-funding-at-1-5b-valuation/ that Nous Research was finalizing a round of at least $75 million led by Robot Ventures, with significant participation from USV, at a $1.5 billion valuation. The financing had not been confirmed as closed. In April 2025, Paradigm led a $50 million Series A https://www.theblock.co/post/352000/paradigm-leads-50-million-usd-round-decentralized-ai-project-nous-research . Wake-word control gives Nous Research another route to make Hermes Agent a persistent interface rather than software users open for isolated tasks. Keeping the listener opt-in and local addresses the first privacy objection, while the broader interaction still depends on how each user configures transcription, models and cloud services.