Quick start

The first time DeskMate starts, it opens the console and walks you through setup. Your first character, Mio, is created for you; she speaks the language your Mac is set to and uses the built-in avatar. The console is also where you add anything later (Models page), so none of these choices are final.

A character needs two things from a model service: a language model to think and talk, and, optionally, a voice to speak aloud. Without a voice the character shows subtitles only.

Pick one path.

Path A: cloud servicesCopy Anchor Link

1. Choose a model. Pick a provider, paste its API key, and choose a model from the list DeskMate fetches. DeskMate talks to Anthropic, Google Gemini and OpenAI directly, and to any OpenAI-compatible service: DeepSeek, Alibaba Model Studio, Kimi, GLM, Doubao, MiniMax, OpenRouter, SiliconFlow, xAI, Mistral, Groq, LM Studio, or a custom endpoint. Before saving, DeskMate sends a short real request and checks whether the model can see images and call tools.
2. Give your character a voice (optional). Voices come from a service: Alibaba Model Studio (CosyVoice), ElevenLabs, MiniMax, Doubao Speech or Fish Audio. Add one, pick a voice, and use Preview to hear it. Skip this step to have the character speak in subtitles.
3. Microphone (optional). Turn it on to talk hands-free. Your speech is recognized locally.
4. Screen reading (optional). Lets the character notice which app you are in and comment on it. See Privacy for exactly what this reads.

Then just talk, or right-click the character on your desktop to type.

Path B: a language model on your own Mac, with OllamaCopy Anchor Link

This keeps the conversation on your machine. It needs no API key, and the character shows subtitles (voices need a vendor).

1. Install Ollama and start it.
2. In step 1 of the onboarding DeskMate detects it. Click Use Ollama. DeskMate picks a model that can see images and call tools for the memory you have (qwen3.5:9b with 16 GB or more, otherwise qwen3.5:4b), pulls it if needed with a progress bar, and creates a derived model with a 16k context window so conversations don’t get cut off.
3. Continue with the voice, microphone and screen steps, or skip them.

If Ollama isn’t installed, the onboarding links to its download page; press the re-detect button once it is running.

Things to tryCopy Anchor Link

• Talk, then interrupt. Speak over the character: it stops at once, and only what you actually heard counts as said.
• Autonomy. In Settings › Autonomy choose how much the character speaks up on its own, from Quiet (only answers) to Autonomous (starts helpful tasks itself). Use Quiet for 30 Minutes in the menu bar when you need peace.
• Memory. The Memory page shows what the character remembers of you, and you can delete any of it.
• Tools. Add MCP servers or allow the character to use specific apps on the Tools page. Anything with lasting consequences (deleting, sending, paying) asks you to click Approve first.
• More characters. The Characters page creates characters, imports 3D avatars (.vrm), and shares or imports characters as .deskmate files.