Huginn
Thought · coding agentA self-hosted, model-agnostic coding agent and harness. It plans a task, edits files, runs the code, reads the failures, and loops until a real test passes — the harness, not the model, doing the heavy lifting.
◆ Sovereign & self-hosted AI
Thought & Memory,
Thought & Memory,
two ravens.
In the old stories, Odin sent two ravens across the world each day — Huginn (Thought) and Muninn (Memory) — and they returned to whisper what they had seen. This is a small flock of AI projects built on that idea: agents and infrastructure that think and remember, running entirely on hardware I own.
Huginn · "Thought"
The thinking raven — a coding agent and harness that plans, edits, runs, and verifies its own work.
Muninn · "Memory"
The remembering raven — a personal assistant that keeps a living memory across every conversation.
The flock
Projects
Everything here is built and run on local hardware — no cloud lock-in, model-agnostic, privacy first. Each one links out to its source. New here? See how it all fits together →
Muninn
Memory · personal agentA personal AI agent you reach over a self-hosted Matrix server, with a persistent, evolving memory. Small context, deep recall: facts are extracted, consolidated, and decayed over time, so it remembers what matters and forgets the noise.
Tyr
Identity · OAuth2 / OIDCA from-scratch OAuth2 / OpenID Connect authorization server — written by hand, not configured from an off-the-shelf product. Not yet wired into the system; it's there to provide sign-in for an in-progress web/mobile interface to Muninn.
Voice Pipeline
Real-time speech infraA wake-word-free, self-hosted real-time speech stack: voice activity detection, streaming transcription, speaker identification, and natural text-to-speech — all on local GPUs, feeding the ravens by voice.
Bifrost
The rainbow bridge · device → homeTurns a spare Android tablet or phone into a first-class Home Assistant device — speaking ESPHome's own native protocol, no webview. One bridge carries it all: BLE proxy, sensors, media player, live camera, a voice assistant, and a fully native audio-reactive visualizer running in per-room sync with the music.
Heimdall
The watchman · Linux voice satelliteWhere the flock's voice began. A wake-word-free Linux voice satellite and synced multi-room speaker — my heavily modified fork of the Open Home Foundation's Linux Voice Assistant, with ambient always-listening and Sendspin multi-room audio added on top.
The idea
Built to be owned, not rented.
These projects share a conviction: capable AI shouldn't require handing your data, your prompts, and your dependence to someone else's servers. Everything in the flock runs on hardware I own and operate — a homelab of GPUs, a Matrix server, a firewall I configure by hand.
They're also a body of work. Each one is a real system that runs every day, not a demo — and together they're the most honest answer I can give to "what can you actually build?"
- Sovereign. Local GPUs, no cloud dependency, your data never leaves.
- Model-agnostic. Any OpenAI-compatible endpoint — swap the brain freely.
- Memory over context. Don't grow the window; grow what's remembered.
- Harness over model. A good loop beats a bigger model on real work.
- Privacy first. Open source, self-hostable, nothing phones home.