Set up AI

Runs LLMs on your own machine. Zero cost, fully offline, your content never leaves your laptop. Best for privacy-first work and when you don't want to think about per-token bills. Heavier on RAM than cloud — a 20B-parameter model wants ~16 GB; a 7B model runs comfortably on 8 GB.

1. Install Ollama.

   macOS / Linux:

   curl -fsSL https://ollama.com/install.sh | sh

   or grab the installer at ollama.com/download.

2. Pull a model.

   Top picks as of April 2026:

   ollama pull qwen3-max:latest      # Alibaba — flagship, top of open-weights leaderboards
   ollama pull qwen3-coder:32b       # Alibaba — best coder you can run locally
   ollama pull gemma4:26b            # Google — strong reasoning, 26B sweet spot
   ollama pull qwen3.5:27b           # Alibaba — fast, multilingual workhorse
   ollama pull glm-4.7-flash:latest  # Zhipu — punches well above its size
   ollama pull deepseek-r1:32b       # DeepSeek — heavyweight reasoning, long thinking

   Ollama auto-starts on port 11434.

3. Point MindOS at it.

   Settings → AI Provider → Ollama. The URL field defaults to http://localhost:11434; change only if you run Ollama elsewhere. Model: whatever you pulled.

Heads up: running a 20B+ model while MindOS is also open can saturate RAM. If the app stutters, pick a smaller model or close other tabs — Ollama holds the model in memory for as long as it keeps getting requests.

Google's API for Gemini. Genuinely free up to a per-day quota (well beyond what a single user sends in a day of writing), and fast. Best balance of "doesn't need a credit card" + "cloud-class quality" right now.

1. Get a key.

   Open aistudio.google.com/apikey — sign in with Google, click Create API key. Pick or create a project when prompted. The key starts with AIza….

2. Paste it into MindOS.

   Settings → AI Provider → AI Studio (Gemini) → API key. The key is encrypted into your OS keychain (Keychain.app on macOS, libsecret on Linux) — not stored in plaintext on disk.

3. Pick a model.

   gemini-3-flash    # default — fast, cheap, 1M-token context
   gemini-3-pro      # frontier reasoning, longer latency, higher quota cost

Google AI Studio rate-limits unpaid keys per minute, not per day in most regions. If you hit a ceiling mid-chat, wait 60 seconds or flip to Ollama.

One key, any frontier model — Claude, GPT, Gemini, Llama, DeepSeek, Qwen, hundreds more. Billed per token on top of the provider's actual price. Best when you want to hop models without juggling five dashboards.

1. Create an account.

   Sign in at openrouter.ai (Google, GitHub, or email). Add credits under Credits — $5 lasts most users weeks of editing.

2. Generate a key.

   openrouter.ai/keys → Create Key. Optional: set a credit limit per key so a bug in one script can't drain the whole balance. Keys start with sk-or-….

3. Plug in.

   Settings → AI Provider → OpenRouter → API key. Pick a model:

   anthropic/claude-sonnet-4.6    # default — best general editor
   anthropic/claude-opus-4.7      # heaviest reasoning, longest context
   openai/gpt-5                   # OpenAI flagship
   openai/o4-mini                 # reasoning, cheap per token
   google/gemini-3-pro            # Google's frontier, 2M-token context
   qwen/qwen3-max                 # Alibaba flagship — tops open-weights charts
   qwen/qwen3-coder-plus          # coding specialist, very cheap
   deepseek/deepseek-v3.1         # strong open-weights, very cheap

   Browse the full list at openrouter.ai/models.

OpenRouter is the right default for "I don't care which model, I want the best available result." Come back to AI Studio when cost matters or Ollama when privacy does.

Which should I pick?

You can switch providers anytime in Settings. Your previous keys stay in the keychain — no re-entry on toggle.