Run your own private AI backend on your Mac or PC, and connect the MyLLM iOS app to it in one tap — over real HTTPS, from anywhere.

MyLLM Connect is a small desktop companion (system-tray app for macOS and Windows) that turns "set up a local LLM server my phone can reach" from a multi-step, HTTPS-and-firewall headache into a single QR scan.

Download

Grab the latest installer from the releases page or from the MyLLM download page. The macOS app is signed and notarized (universal — Apple Silicon + Intel). The Windows installer is not code-signed yet, so SmartScreen will warn — signing is on the roadmap. You'll also need Ollama and a free Tailscale account; the app guides you through both on first run.

It does three things:

1. Runs a local model server on your machine (manages Ollama for you).
2. Gives it a trusted HTTPS address your iPhone accepts — privately, over your own Tailscale mesh, with a valid certificate. No port-forwarding, no self-signed-cert warnings, works on your home WiFi and when you're out.
3. Pairs to MyLLM with a QR code — scan once and the app is configured with the address and a private access key.

Your prompts go straight from your phone to your own machine. Nothing runs in our cloud; we never see your data.

Why HTTPS matters

iOS only trusts a valid certificate. A plain http://192.168.x.x:11434 Ollama server is rejected by the app's transport security — which is why "just point the app at my PC" usually fails today. MyLLM Connect solves this by giving your machine a real, trusted HTTPS endpoint automatically.

Locked to MyLLM

The endpoint is useless without the access key minted during pairing, and that key only lives inside your paired MyLLM app. The companion is free and open; the experience it unlocks is the MyLLM app. (See PAIRING_PROTOCOL.md.)

Start here (server team)

This repo is the whole handover — no separate doc to chase.

1. Read the pinned Epic #11 — it sequences every issue, gives the build order, and the definition of done.
2. Read SPEC.md (what to build, recommendations, open questions) and PAIRING_PROTOCOL.md (the exact contract).
3. First moves: close #1 (ADR-001 runtime) and #2 (ADR-002 model server) — record them in docs/ARCHITECTURE_DECISIONS.md (ADR-003 HTTPS=Tailscale is already decided) — and post your estimate on #10.

You can test against the live app today

The iOS half is already built and shipped (MyLLM v2.6 is on the App Store). You do not need to wait for the full companion to validate pairing:

• In the app: Settings → Server → Pair a Backend (scan QR).
• Build a QR encoding myllm://pair?v=1&url=<https>&token=<key>&models=<id> per PAIRING_PROTOCOL.md and scan it.
• Stand up the endpoint by hand first — ollama serve behind a bearer-checking proxy, exposed with tailscale serve — and you can prove the entire scan → chat-over-HTTPS loop before writing the tray app. Build #3 (auth proxy) and #4 (Tailscale) and you're already testable end-to-end.

Status

Spec, pairing protocol, and ADRs are in place; implementation is tracked in the issues (start at the pinned epic). The iOS pairing flow it targets is already shipped in MyLLM v2.6.

The Windows companion now implements the full pairing flow (in-process auth proxy → Tailscale serve → QR), verified end-to-end against the live app. The Rust core is cross-platform; macOS team: see docs/MACOS_IMPLEMENTATION.md for the macOS-specific punch list (Tailscale CLI path, menubar/template icon, signing/notarization, flipping on the CI .dmg job).

Sharing beyond yourself: trusted circles

This is the personal backend onramp: your phone, your server, your data. The next step on the roadmap is trusted circles — inviting a few people you actually know (your household, a small team) to pair their MyLLM apps to your machine, each with their own key and their own QR code, individually revocable.

An open federation — sharing your LLM with strangers — is deliberately not the plan. A privacy-first app shouldn't route anyone's prompts through hardware they have no reason to trust, and ordinary consumer machines can't prove they aren't logging what passes through them. Circles keep the trust model honest: you share with people who trust you.

License

Apache-2.0.