nixai — AI Agent Host

Purpose

nixai is the AI infrastructure host. It runs the AI agent, local inference via Ollama, the Open WebUI chat interface, Mnemosyne memory sync, and voice services.

Entry Point

• Main file: hosts/server/nixai/default.nix
• Supporting files: ai-agent.nix, backend.nix, mnemosyne.nix, voice.nix, web.nix

Architecture / Services / Scope

AI Agent

• Containerized AI agent with a web dashboard, an OpenAI-compatible API server, and a webhook platform.
• Memory is enabled, backed by Mnemosyne sync.
• Voice is enabled, using the local Wyoming STT server.
• A hardened SSH daemon is provisioned inside the container for agent administration.

Local Inference

• Runs local models on the host’s AMD iGPU (ROCm/Vulkan), loading a small set of chat and embedding models.
• Used by Open WebUI and the agent for local model access.

Open WebUI (web.nix)

• User-facing chat interface with RAG, web search (via SearXNG), and local Whisper STT.
• Backed by a PostgreSQL database on the Database Coordinator and tied into the database availability target so it only starts when the DB is reachable.

Mnemosyne (mnemosyne.nix)

• Mnemosyne memory provider sync server, used by the agent for persistent memory. Served on its own vhost.

Voice (voice.nix)

• Wyoming Piper (TTS) and faster-whisper (STT) servers exposed over TCP for local voice assistants.

Secrets

Declared secrets

Operational Notes / Assumptions

• Ollama relies on the AMD iGPU; ROCm on the iGPU is noted as unreliable upstream, so the config pins an override and waits on a Vulkan fix.
• Open WebUI, the agent dashboard, Mnemosyne, and the voice/whisper endpoints are exposed through the IO Coordinator reverse proxy.
• The agent’s SSH port is opened to the subnet for administration.

References

• Hermes Agent
• Open WebUI
• Ollama
• Mnemosyne
• Database Coordinator
• IO Coordinator