A private AI agent that runs on the Mac you already own.

Cells, not conversations

Small on-device models lose the thread in long chats, so Swiftea never has one. Each unit of work is a cell: one short, bounded model session with typed input and output and at most four tools.

Cells share a small written state object instead of a growing history, and deterministic Swift code walks the task graph (triage, locate, digest, plan, edit, verify, report) deciding what runs next. Files too big for the model's window are split, digested separately and recombined in Swift.

Safety built in, not bolted on

The agent's six tools sit behind a four-layer permission gate. Shell commands are checked against a hard denylist and run inside a kernel-enforced macOS sandbox; the shell tool refuses to start at all without one. Every edit is snapshotted so a task can be rewound.

Measured, not claimed

On September 21, 2026 Swiftea fixed a seeded bug end to end on Apple's on-device model, with no network and no cost. Runs are scored on ground truth: each task's own tests are re-run afterwards, and what the agent says it did is never counted. Some eval tasks are traps where the right answer is to refuse.

A native window into every run

The SwiftUI app shows each run as a cell timeline with tokens, timings and how much ran on-device, plus a diagnosis of where a run stopped. Accessibility is a release gate: text scaling from 85 to 200 percent, VoiceOver labels and reduced-motion support.