takemo101’s voice-control uses Tiny-Jev to select a Mac automation skill, then a language model to carry it out. The Swift prototype includes inspectable inference code, command restrictions and unusually detailed records of failed experiments.

Give each component a separate job

Tiny-Jev is a downloadable decision model used locally here, separate from TypeSafe’s hosted Jev service. The implementation loads pinned weights and scores candidate descriptions through Apple’s GPU framework.

  • Skills declare their instructions and permitted command prefixes in SKILL.md.
  • Tiny-Jev chooses among enabled skill descriptions, including a “none” option.
  • Apple’s on-device language model, or an optional local server, proposes commands through run_command.
  • Code checks each command against the selected skill’s allowlist before execution, without a shell.

The fallback skill can operate apps through agent-desktop. Its decision about whether a request needs confirmation still depends on a model judgment. Named skills use their confirmation setting and selection confidence.

Spoken approval exposed a second classification problem

The author reports that Tiny-Jev accepted only 3 of 16 affirmative Japanese replies under the original confirmation rule. Adding confirmation context fixed those examples but also made unrelated responses such as “thank you” count as approval.

The current code accepts a controlled set of affirmative phrases directly. Other replies go through Tiny-Jev with a confidence cutoff. This separates common approval words from a broader semantic guess.

The audio layer matters too. In the author’s synthetic silence-and-noise test, Apple’s recognizer produced “yes” on 6 of 20 recordings containing no speech. The project added a speech-presence check before transcription. Its thresholds were tuned on the same synthetic material; real-room reliability remains unmeasured.

Keep date ordering and permissions in code

The author reports Tiny-Jev selecting the intended file in 14 of 16 constructed scenarios. Comparing “newer” and “older” caused failures. Adding today’s date reduced the result to 6 of 16, so the implementation leaves ordering to code.

The file-search skill combines Spotlight search, optional Tiny-Jev selection and opening a file. Tests used constructed folders and a substitute open command. The documented results include adjustments made on the same examples.

  • Command permission is checked separately from selection probability.
  • A permitted prefix still allows its trailing arguments; broad commands carry broad authority.
  • A typed answer checks shape. Correct selection, valid consent and the intended application effect each need their own evidence.

Read the failure log alongside the demo

The verification record distinguishes current features from a retired direct-control experiment. It also records abandoned screen-reading approaches, prompt changes and remaining gaps. README images show real interface components with staged content.

The repository requires Apple silicon and macOS 26 or later. Its code is MIT-licensed; the author explicitly leaves the model’s license terms unverified. Confidence thresholds and the destructive-action classifier still need broader evaluation.

Verification: inspected the README, failure records, local inference, routing, approval and command-policy code. The macOS application and author-reported measurements were not run or reproduced.