Desert Ant CLI

Transcribe a recording, cut the best moments into clips, clean up the audio, redact personal data from text, name the language of a line, flag nudity in a photo, or tag a topic, with Desert Ant Labs models running on your own machine. Weights download once. Nothing leaves the device.

$ da voz talk.mp4 -t
0:00  We build small models that run on the device.
0:04  Nothing leaves the phone, and there is no cloud bill.
Transcribed 48 minutes of audio in 9.7s, 296x realtime.

$ da clips talk.mp4
 1  10:14 to 10:42  27s
    The one thing the meter can't sell you is inference that costs nothing.
    talk_clip-1.mp4
8 clips from 48 minutes in 41s.

$ da redact "Email Anna at [email protected] or call 555-0100"
Email [GIVEN_NAME_1] at [EMAIL_1] or call [PHONE_1]

da is short for desertant.

curl -fsSL https://raw.githubusercontent.com/Desert-Ant-Labs/desert-ant-cli/main/install.sh | sh

Or brew install desert-ant-labs/tap/desertant, or mise use -g "ubi:Desert-Ant-Labs/desert-ant-cli[exe=desertant]".

Each model has a page on desertant.com with its numbers, its limits, and a live demo. da info <model> prints the same card.

• Voz da voz talk.mp4
  Transcribe a recording, with a time on every word, in 25 languages.
• Clips da clips talk.mp4
  Cut a podcast, interview, or meeting into short clips and highlights.
• Clear da clear talk.mp4
  Studio sound: noise and room echo out, loudness leveled.
• Uhm da uhm talk.mp4
  Find every "uh", "um", and "hmm", with the exact span to cut.
• Ear da ear talk.mp4
  Name the language being spoken, from 99.
• Redact da redact "<text>"
  Replace names, emails, and phone numbers in a text with labels.
• Emo da emo "<text>"
  Suggest emoji for a word or a sentence.
• Gist da gist "<text>"
  Tag a post or article with topics, from a set of 36.
• Tongue da tongue "<text>"
  Name the language a short text is written in, from three words, in 84 languages.
• Moderator da moderator photo.jpg
  Flag nudity or sexual content before a photo is uploaded or shown.
• Title da title "<text>"
  Suggest a short factual title and a one- or two-sentence description for any text.

Redact, Emo, Gist, and Tongue are multilingual. Voz, Uhm, Title, and Moderator need Apple silicon, the rest run on macOS and Linux, and da models lists what runs on your machine.

da voz talk.mp4 prints the transcript, one sentence per line, and -t puts the time in front of each. Caption and text files are one flag each, written beside the input:

da voz talk.mp4 --srt          talk.srt, captions
da voz talk.mp4 --vtt --txt    talk.vtt and talk.txt
da voz talk.mp4 -o notes.vtt   one file, the format from its extension
da voz talk.mp4 --format srt   SRT on stdout, for a pipe

Captions are cut the way broadcasters cut them: two lines of 42 characters, no more than 7s on screen. da docs voz has the rest.

Commands compose through their JSON. Transcribe once, then cut:

da voz talk.mp4 --json | da clips talk.mp4 --transcript -

da schema --json names which command's output feeds which, and da docs pipelines covers the rest. desertant never overwrites a file: a second run writes talk_clear-2.mp4, and --force is the only way around that.

A file verb takes several files, or a folder, and loads the model once for all of them:

da moderator photos/ --json
da ear interviews/ --recursive

Under --json the result is an array with one entry per file. A file that fails is reported on stderr and the rest still run. The exit code is 1 at the end.

One command writes a skill for every agent installed on your machine, in the current project: Claude Code and Pi get a skill file, Codex gets a section in AGENTS.md.

desertant setup

desertant setup claude does one agent, and --global writes where the agent looks for every project. Or paste this into the agent and let the agent do both steps:

Install the Desert Ant CLI with: curl -fsSL https://raw.githubusercontent.com/Desert-Ant-Labs/desert-ant-cli/main/install.sh | sh
Then run `desertant setup` in this project.

Any other agent with a shell needs nothing more: every command returns JSON with --json, reads stdin, and exits 0, 1, or 64, and da schema --json describes each command and how they chain.

da docs prints the docs offline. The same docs are in the repo, indexed by llms.txt: using desertant from an agent, the JSON contracts, files and the cache, clips, pipelines, and performance.

MIT for the CLI. The models desertant runs are under the Desert Ant Labs source-available license.