Hermes, OpenClaw and NanoClaw propose GPT-Live voice fronts for their existing agents

Hermes, OpenClaw and NanoClaw have opened proposals that separate live conversation from agent execution. OpenAI's GPT-Live handles the audio exchange; requests needing memory, tools or actions are delegated to the existing agent. The new integration work follows the public Live API described in OpenAI's current guide, rather than simply attaching another transcription button to a chat window.

The facts

- Hermes PR #108137 proposes an opt-in Desktop voice engine; its existing speech-to-text, agent-turn and text-to-speech chain remains the default. - The Hermes proposal keeps the OpenAI key server-side and delegates requests through the selected Hermes model, tools, memory and approvals. - OpenClaw PR #145382 distinguishes Platform-key gpt-live-1 sessions from the ChatGPT subscription route and preserves explicit model pins. - NanoClaw PR #3764 supplies an add-voice skill, while PR #3772 supplies its adapter on the channels branch; the payload must land first. - NanoClaw's proposal treats the call-link token as caller authority and explicitly leaves stronger caller identity, cost caps and OneCLI key routing out of scope. - The NanoClaw adapter review lists unresolved problems including session creation before access admission and delivery reporting success when nothing was spoken.

Why it matters

The consequential choice is who owns the action, not whose voice sounds most natural. A separate voice frontend can preserve a familiar agent's tools and policies, but it introduces another session lifecycle: interrupting speech, stopping backend work and finishing transcript persistence are different events. OpenClaw explicitly assigns transcript persistence to its Gateway. Hermes describes a newer delegation stopping the in-flight turn through its existing stop path. Those distinctions matter when a spoken request can change files or contact another service.

Current

Inspected on 2026-09-12. The Hermes Agent release baseline is v2026.9.11 published 2026-09-11T19:20:31Z. The main source was open (not merged) when captured. Proposals, merged mainline changes and packaged releases are different availability states; the linked release baseline does not imply that open proposals have shipped.

Evidence

Primary evidence: NousResearch/hermes-agent PR #108137 (https://github.com/NousResearch/hermes-agent/pull/108137). Supporting context: OpenAI — current GPT-Live delegation guide (https://developers.openai.com/api/docs/guides/live); openclaw/openclaw PR #145382 (https://github.com/openclaw/openclaw/pull/145382); nanocoai/nanoclaw PR #3764 (https://github.com/nanocoai/nanoclaw/pull/3764); nanocoai/nanoclaw PR #3772 (https://github.com/nanocoai/nanoclaw/pull/3772). Sources were read directly through public pages or GitHub API bodies. Contributor test results are attributed claims, not independently reproduced results.

Operator take

Treat these as integration designs to compare, not three installation recommendations. A useful trial would verify the selected account, a harmless delegated tool call, interruption, final transcript retention and a confirmed hangup. NanoClaw's disclosed adapter defects are reasons to wait for the follow-up review, particularly before sharing a call link. Hermes reports synthetic-microphone Desktop tests and OpenClaw reports isolated authenticated probes; those are contributor receipts, not tests performed for this report. They establish what each author says was exercised without certifying every device, deployment or model route.

Caveat

All three integration proposals were open and unmerged in the captured snapshot. OpenAI's guide establishes the API contract, not downstream readiness. NanoClaw also says a fresh-clone copy path cannot yet be verified because its payload has not landed upstream; do not expose that proposed voice route as a production service.