Hermes sketches remote voice and desktop-control interfaces

The voice proposal makes a client deliberately thin: send microphone PCM, receive speech PCM and consume typed lifecycle events. Utterance detection and incremental sentence synthesis live on the server. A single-clock device can request matching input and output rates, while browser clients can keep the split rates. Quiet notifications measure received audio silence, not idle wall time while the agent is thinking or speaking.

The remote computer-use proposal points in the opposite direction. A headless agent dials out to a machine with a real display and driver. Availability is determined by the configured remote runtime, not by whether the agent host has a local GUI binary. The bridge enforces authenticated transport and avoids passing its token into the child process environment. A configured remote target that cannot be reached is a failure, not permission to control the local desktop instead.

The facts

- PR #103279 proposes WS /v1/audio/converse, with the server owning speech detection, transcription, the agent turn, synthesis and interruption. - The voice proposal carries authentication in a WebSocket subprotocol or first start frame, not a URL query parameter. - Input and output sample rates default to 16kHz and 24kHz and are bounded between 8kHz and 48kHz. - PR #103653 proposes a remote computer-use provider whose host owns the display, driver and permissions; remote failure must not launch a local driver. - PR #101052 proposes profile-scoped AgentRuntime registration while Hermes retains tool execution, approval, transcripts, delegation and cancellation.

Why it matters

Together these designs separate where intelligence runs from where a person speaks or an application is operated. That can make a small device a useful agent interface without installing the whole agent stack there. It also creates new network and identity seams. Voice authentication must not leak into access logs, and desktop control must identify the machine whose screen and permissions are being used. ‘Remote’ is a routing decision with authority attached, not a cosmetic provider name.

Current

Inspected on 2026-09-06. The Hermes Agent stable-release baseline is v2026.8.31 published 2026-08-31T19:29:49Z. The main source was open when captured. Release metadata is a version boundary, not evidence that an open proposal has shipped.

Evidence

The primary source is NousResearch/hermes-agent PR #103279 (https://github.com/NousResearch/hermes-agent/pull/103279). Supporting context comes from NousResearch/hermes-agent PR #103653 (https://github.com/NousResearch/hermes-agent/pull/103653); NousResearch/hermes-agent PR #101052 (https://github.com/NousResearch/hermes-agent/pull/101052). The linked records were inspected directly; related project records are not independent confirmations.

Operator take

Review these as three related design proposals, not one shipped bundle. The runtime-plugin proposal supplies a useful boundary: an external whole-turn transport does not get to mint trusted transcript rows or bypass Hermes’s approval and effect path. Validate interruption, unavailable remote targets, reconnect behavior and exact profile selection before treating such interfaces as unattended infrastructure. Keep endpoint credentials out of URLs and ensure a failed remote bridge cannot silently substitute a different machine.

Caveat

The voice, remote-control and runtime-interface records were open. The computer-use change is stacked on another provider seam, and the runtime proposal explicitly awaits a maintainer decision. No new platform support or authentication policy is inferred from these proposals.