MyOwnLLM

An AI that runs on every device you own — and manages itself.

Download · Docs · Architecture · Contributing

Install

# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/mrjeeves/MyOwnLLM/main/scripts/install.sh | sh

# Windows
irm https://raw.githubusercontent.com/mrjeeves/MyOwnLLM/main/scripts/install.ps1 | iex

Signed installers for every platform: myownllm.net.

One binary, three personas

myownllm          # desktop GUI (Tauri + Svelte 5)
myownllm serve    # OpenAI / Ollama / Anthropic HTTP on :1473
myownllm run      # terminal chat

What it does

• One AI across your devices. Boxes on the same Network ID find each other via the bundled myownmesh daemon — signed signaling over public Nostr relays, end-to-end WebRTC data channels between peers. No MyOwnLLM-operated broker, no account, no API key. Share inference, pass conversations, pin a peer per surface. Self-host a relay for an air-gapped LAN.
• Every device makes it stronger. A Pi can borrow the workstation's model over the mesh; the workstation can serve the laptop's chat. Manifest-driven hardware tier selection covers Pi 4 2 GB through a 4090; edge variants for Gemma 4 (e2b/e4b) and Qwen 3.6 down to 0.8b.
• It's its own IT. The chat is a tool-calling agent with four tools today — networks (mesh: status / add / approve / switch / forget / reconnect / rediscover / accepting / recent activity), shell, read_file, write_file. shell and write_file route through a per-network permission gate (Deny / Allow once / Always for this command-or-path / Always for the tool); decisions persist per network in Config.cloud_mesh.networks[*].agent_permissions and gossip to peers on the same network when auto_gossip is on. Tools execute on the caller's box even when inference is on a remote peer — your Pi can borrow the 4090 and still configure the Pi.
• Diarized live transcription. Moonshine or Parakeet TDT (25 languages), pyannote-segmentation-3.0 + online clustering. Speaker IDs stay stable across the whole session — not per window. In-process; no Python venv, no whisper sidecar.
• Talking Points. A live LLM loop summarises the growing transcript into a bullet list while you talk. Pausable, persisted with the session, on-device.
• Three wire formats, one server. OpenAI on :1473, plus Ollama and Anthropic. Cursor, Continue, Aider, Cline, Zed, Open WebUI, opencode all just work.
• Self-updating. Stages on launch, applies next start. Last-good manifest cached for offline runs.

Drop-in OpenAI

myownllm serve &
curl http://127.0.0.1:1473/v1/chat/completions \
  -H 'Authorization: Bearer myownllm' \
  -d '{"model":"myownllm","messages":[{"role":"user","content":"hi"}]}'

Platforms

macOS 12+ (Apple Silicon) · Linux x86_64 / aarch64 · Windows 10+ · Raspberry Pi 4 & 5. Signed, auto-updating, ~50 MB app plus first-run model.

More

DOCS.md — manifests, CLI, API, lifecycle, scripting · ARCHITECTURE.md — internals, modules, data flow · CONTRIBUTING.md — setup, repo layout, commit style · LICENSE — MIT