Cloud AI assistants vs. local-first: what actually runs where.

Every AI assistant says it's "private." Almost none tell you what runs where. Here's the honest version.

The three architectures

• Pure cloud (ChatGPT, Copilot, Gemini). Everything — your words, your files, your usage patterns — lives on the provider's servers, governed by their policies, which change. You are a tenant.
• Pure local (Ollama + open models). Everything on your machine. Genuinely private — but you trade away frontier-model quality, and you're assembling plumbing, not using a product. You are a hobbyist. Respect.
• Local-first with your own keys (Selam's lane). The application — interface, avatar, speech recognition, camera awareness, memory, files — runs on your Mac. Only the model calls leave, going directly from your machine to the AI provider you chose, on your API key. The vendor in the middle (us) never sees the traffic. You are an owner.

What "local" means in Selam, concretely

• Speech-to-text: local — faster-whisper running on your Mac
• The avatar and lip-sync: local — Three.js, rendered on-device
• Camera awareness (she pauses her presentation when you step away): local — MediaPipe on-device; no frame ever leaves the machine
• Your agent's memory: local — plain Markdown files, one-click export
• API keys: local — macOS Keychain
• Model inference: your chosen provider, direct — Anthropic, OpenAI, Gemini, Grok, Mistral, Kimi, DeepSeek
• What touches our servers: license validation and update checks. That's the list. Full security model →

The cost math nobody shows

Similar totals by year three — but the composition matters. With Selam the marginal dollars buy raw model usage you control (hard caps included), the license never renews, and when model prices drop you benefit — not us. Stop paying a subscription and you lose everything; stop using Selam and you keep the software and every memory file.

When cloud is genuinely right

You want mobile, zero setup, or your company standardized on one. No shame in it — but call it renting, because that's what it is.

When local-first is right

You talk to your assistant about things that are nobody else's business. You want frontier models without marrying one vendor. You want an agent that accumulates years of context you own.