[raullenchai]

What changed

Adds docs/ESP32-THINKS.md and links it from docs/README.md.

The page describes how an ESP32 voice gadget (esp32-thinks) can sit in front of OpenMuse. The device speaks OpenAI-compatible STT/chat/TTS; OpenMuse's agent transport is AG-UI, so a server-side gateway is required. The page documents that gateway contract, deployment/TLS requirements, and the security boundary.

Scope

Docs-only. No code, tests, or runtime changes. It follows the same integration-boundary pattern as docs/OPENBOT-INTEGRATION.md.

Verification

• Links checked locally.
• The referenced firmware and its OpenAI-compatible gateway are tested on a Seeed SenseCAP Watcher; the AG-UI bridge to OpenMuse is explicitly marked as not implemented or tested.

Notes

Happy to move this under a different docs path or turn it into an examples/ entry if you prefer.

[jerelvelarde]

Reviewed 3095eb3eef800a415abc6245dfc1be8f05b4929e.

Value/template fit: this is a bounded documentation-only integration sketch, clearly distinguishing the external OpenAI-compatible gateway from OpenMuse's AG-UI runtime. It follows the existing OpenBot boundary-document pattern and explicitly states that the bridge is not implemented or tested. Keep this open for merge rather than closing it as if it were a runtime feature.

Security review: checked the proposed bearer-auth gateway boundary, server-side OpenMuse credential handling, thread continuity, cancellation, honest failures, and the explicit rule that the device cannot approve writes or grant tool permissions. No executable code, dependency or permission changes, and no actionable runtime security finding. For implementation, clarify that direct device-to-TTS authentication needs a scoped device credential or a gateway proxy; the statement that provider credentials stay server-side should not be read as automatically true for the diagram's direct TTS leg. Derive device/thread authorization from the authenticated identity, not an untrusted hardware ID alone.

Verification: read the entire diff, checked local relative references and the existing integration precedent, and cross-checked the linked firmware project's docs/OPENMUSE.md, which likewise distinguishes its gateway from CopilotKit/OpenMuse. No hardware, TLS-edge behavior, voice transport or live AG-UI bridge was exercised; reported device behavior remains contributor evidence. CI was approved and merge is gated on completion of all seven jobs. No code blocker found.