// the take
FlowDown is a native Swift/UIKit AI chat client for iOS, iPadOS, and macOS that lets you plug in cloud models (via any OpenAI-compatible API), self-hosted servers like Ollama/LM Studio, or on-device MLX models. It's aimed at power users who want to configure and run their own model setups rather than a one-click consumer chatbot.
Privacy is actually enforced in the architecture, not just promised: no telemetry or crash reporting, all state stored locally or in the user's own iCloud, and the client code is open so the claim is checkable. It ships a genuinely unusual built-in evaluation harness (EvaluationManifest/EvaluationSession with per-language and tool-calling test suites) that lets you benchmark whatever model you've wired up instead of trusting someone else's leaderboard. It's also more than a chat wrapper — MCP client support plus native tool integrations for Calendar, Reminders, Location, memory, and web scraping make it a real agent host. Using Swift/UIKit with MLX instead of a web-wrapper shell means it isn't paying Electron-style overhead for UI or inference.
It's Apple-only — nothing for Windows, Linux, or Android, so it's a non-starter outside that ecosystem. The code is AGPL-3.0, but the name, icon, and artwork are proprietary and commercial use needs a separate license from the team, so this isn't a clean fork-and-rebrand situation despite the open-source label. There's no server or multi-user layer at all: it's a single-user client, so anything beyond personal use is out of scope. And distribution leans on the App Store with an in-progress (not yet finalized) move to a free-plus-subscription model, so building and maintaining it yourself means tracking a fast-moving Xcode project with a non-trivial SPM dependency graph rather than installing a stable package.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →