LLMster

LLMster connects straight to your LM Studio server and streams chat, reasoning, tool calls and vision — in a fast, private app that feels like it shipped with iOS.

Nothing leaves the network LLMster talks only to the LM Studio server you point it at. No telemetry, no accounts, no third-party cloud in the path.

No account, no copies There's no LLMster sign-in and no copy of your conversations passing through third-party servers. Your history stays between your device and the server you chose.

On-device speech Voice chat uses on-device speech recognition. Talk to a model that's running on the machine in the next room — and nowhere else.

No desktop app required. Keep your models on the machine that can handle them — a server in the closet or a workstation that never sleeps — and reach them from your phone. Run LM Studio, or just the standalone llmster daemon. Point the app at the host and start chatting.

The right model for the moment. A fast 4B for quick questions, a 30B that reasons, a vision model for screenshots — switch between them in a tap, without ever leaving the conversation.

Find it fast — search or sort by size so the model you want is one tap away, even with dozens installed. Know before you load — see at a glance which models can see images, reason, or call your tools. Swap mid-chat — change models without losing the thread; pick up right where you left off.

Show it what you mean. Attach an image and ask. LLMster sends it to any vision-capable model on your server and renders the answer right in the thread — pictures are downscaled and kept on your device.

Up to four images — drop in screenshots, photos or scanned pages per message. Zoom right in — open any image full-screen to read the small print or check a detail before you ask. Never leaves your device — images are resized on your phone, so the originals stay with you. Any vision model — routed to whichever vision-capable model you've loaded.

Talk to your model. Out loud. Start a hands-free conversation and trade turns with your local model in real time. LLMster listens with on-device speech recognition and speaks the reply back with on-device text-to-speech — the whole exchange stays on your hardware.

On-device STT — your voice is transcribed locally and sent as text. No audio leaves the device. On-device TTS — replies are spoken back with the system voice, no cloud speech service in the path. Natural turn-taking — it listens, you speak, it answers, then waits for you again — fully hands-free. Haptic cues — a subtle tap when your turn is sent and when the reply lands.

Tune every conversation. Every chat carries its own settings. Switch integrations on for the task at hand, give the model a role, and decide exactly how much of its work you want to see.

MCP integrations — flip servers like Linear, Notion or Figma on per chat. Custom system prompt — set the model's role; it's sent with every later message. Show or hide tool calls — see exactly what the model did, or keep the conversation clean and conclusions-only. Reading, your way — hide the model's thinking and keep the view steady while it writes, so nothing jumps around as you read.

Know the moment it's done. Kick off a long reply, then go do something else. LLMster tracks it from the Dynamic Island and taps you on the shoulder the second your model finishes — fired right on your phone, with no push servers in between.

Local-only alerts — notifications come from your phone, not through a push cloud that sees your chats. Live in the Dynamic Island — glance up to see it thinking, then receiving, without opening the app. Lands on your lock screen — the finished reply is waiting for you, ready to read at a tap.