What is HiveMind?

HiveMind is an open-source protocol and platform that connects lightweight satellite devices to a central AI hub over the network — including hardware that cannot run a full AI system on its own.

The core idea

HiveMind separates AI workload from edge devices. Satellites (microphones, voice devices, browsers, chat clients) delegate processing to a hub, which handles reasoning, skills, and audio processing. The hub can serve many satellites simultaneously with independent sessions and permissions for each.

[Mic Satellite]  ──┐
[Voice Relay]    ───┤──→ [HiveMind Hub] ──→ skills / LLM / intents
[Voice Satellite]──┤         │
[CLI Client]     ───┘        └──→ responses back to each satellite

Hub types

Hub Package Agent backend
OVOS skills server hivemind-core OpenVoiceOS (full skill ecosystem)
Audio hub hivemind-core + hivemind-audio-binary-protocol OVOS + server-side STT/TTS/wakeword
Persona server hivemind-persona LLMs and chatbots via ovos-persona

Satellite types

Satellites form a spectrum based on where audio processing happens. At one end the satellite does nothing but pass text; at the other it handles the complete voice stack locally.

Satellite Local processing Hub requirements
HiveMind-cli Text I/O only Any hub
hivemind-mic-satellite Mic + VAD Audio binary protocol for STT/TTS/wakeword
HiveMind-voice-relay Mic + VAD + Wakeword Audio binary protocol for STT/TTS
HiveMind-voice-sat Mic + VAD + Wakeword + STT + TTS Any hub (sends text utterances)
hivemind-webspeech Browser mic + VAD Any hub

See Choosing a Satellite for a detailed comparison.

Mesh topology

HiveMind hubs can connect to other hubs, forming a hierarchy. A sub-hub acts as a client to its parent while being a server to its own satellites. ESCALATE messages travel up the chain; BROADCAST messages travel down. This enables multi-room or multi-household topologies where each hub controls its local devices but can route requests upward.

[Master Hub]
    ├──→ [Sub-hub A] ──→ satellites
    └──→ [Sub-hub B] ──→ satellites

See Nested Hives for details.

Security model

Every client authenticates with an access key and a password. After authentication the session is encrypted with AES-256-GCM. The key is never transmitted; both sides derive it independently via PBKDF2 from the shared password and exchanged nonces.

Permissions are configured per client. The default policy admits only a basic set of message types; you expand or restrict per-client using the hivemind-core CLI. Skills and intents can be blacklisted individually.

See Security for the full model.

What HiveMind is not

  • It is not a replacement for OVOS when running as a skills server. OVOS remains the AI back-end; HiveMind is the external-facing gateway.
  • It is not a cloud service. Everything runs on your own hardware.
  • It does not replace the OVOS messagebus for local communication. The messagebus remains internal.