One command instead of a throwaway script.

One binary for the file jobs agents otherwise script. In our benchmark, Codex used 27% fewer tokens and 57% less time with it.

Install deft How it was measured

This is how an agent finds a command. Try it: the search below uses deft's real index.

$ deft find

An agent doing the job.

Video slot: the demo goes here. Add site/media/demo.mp4 before hosting.

Same task, same agent.

Recorded results from the benchmark. The last task is one deft lost, because a plain tar is hard to beat.

Measure	With deft	Without

Fewer tokens, less time, same work.

100 tasks written the way a person asks. Totals cover the 85 that deft supports.

85/85tasks solved with deft, 78 without

3 vs 70failed commands, with deft and without

27%fewer tokens, counting uncached input plus output

57%less time, wall clock per task

Under 10 msfor a help answer, measured on the release binary

50 MBbinary, about; ffmpeg and yt-dlp are separate

How it was run

Agent: Codex with gpt-6-luna at medium effort, in its workspace-write sandbox with no network.
With deft: The same setup, plus deft on the PATH and its generated skill text as AGENTS.md.
Without: Whatever the machine has. That included ffmpeg, Ghostscript, Pillow and pandas, so the baseline is stronger than a clean machine.
Judging: Checks run with Ghostscript, ffprobe, Pillow and openpyxl, not with deft's own libraries.
Left out: 9 tasks deft cannot do yet and 6 that need the network, which were not run.

Read this before quoting the numbers. The latest attempt counts: 22 of 188 task and arm pairs were rerun, and 8 of those went from fail to pass, 4 in each arm. One agent; Claude Code has not been measured. Deft was improved against these same tasks, so this is a development benchmark and not a held-out test.

The cases, results and harness are in evals/.

Where it helps, and where it does not

Savings come from jobs where an agent would write code. On trivial jobs deft can cost more, because the agent reads help first: it used more tokens on 31 of the 85 tasks.

Group	Tasks	Tokens	Time
convert	3	63% less tokens	80% less time
pdf	20	42% less tokens	68% less time
md	4	36% less tokens	87% less time
image	18	35% less tokens	53% less time
sheet	9	31% less tokens	53% less time
video	16	7% less tokens	5% less time
audio	6	19% more tokens	2% more time
subtitle	5	33% more tokens	11% less time
archive	4	57% more tokens	44% more time

173 commands in 17 groups.

This list is generated from the source. One syntax for all of it: deft <group> <command> [inputs] [-o output]. Full table in docs/COMMANDS.md.

pdf

Merge, split, compress, encrypt, forms, OCR, compare, convert

infoPages, size, metadata, fonts, images
compareCompare two PDFs: text and pixels
compressMake smaller: --level, --target-size 2MB
cropTrim pages: --margin, --box, --auto
decryptRemove a password you know
deleteRemove pages: --pages 2,5-7
encryptProtect with a password (AES-256)
extract-imagesSave embedded pictures, unchanged
fieldsList form fields and their values
fillFill a form from JSON: --data values.json
flattenBake forms and annotations into pages
from-imagesOne page per image
from-mdTypeset Markdown as PDF
grayscaleRemove colour, keeping text as text (Ghostscript)
mergeJoin PDFs in the given order
metaRead metadata and bookmarks, or set --title --author
numberAdd page numbers: --format
ocrMake a scan searchable
redactBlack out for good: --box or --text
repairRebuild a damaged PDF (Ghostscript)
resizeFit pages to --page-size a4
rotateTurn pages clockwise: --angle 90, --pages
searchFind text: --text invoice
selectKeep and reorder pages
splitSplit: per page, --every N, --ranges
stripRemove metadata
to-docxEditable Word file, layout not kept
to-imagesRender pages: --dpi, --pages
to-mdText as Markdown paragraphs
to-pdfaConvert for archiving as PDF/A-2 (Ghostscript)
to-sheetPull tables out by text position
to-textExtract text
watermarkStamp text or an image on pages

image

Resize, crop, convert, compress, filters, montage, OCR

infoFormat, dimensions, EXIF orientation, GPS
adjustChange --brightness, --contrast, --saturation, --gamma
blurGaussian blur: --radius 5, or a --box region
compressSmaller file: --quality, --target-size 200KB
convertChange format: --to png|jpg|webp|avif|gif|tiff|bmp|ico
cropCut a region: --box x,y,w,h or --aspect 16:9
diffCompare two images; marks changed pixels in red
faviconWebsite icon set: favicon.ico and PNG sizes
flipMirror: --direction horizontal|vertical
framesSave each frame of an animated GIF as PNG
grayscaleRemove colour
invertNegative: flip every colour
metaRead EXIF, or set --title --author --copyright --date
montageTile images into a grid: --cols 3 --size 300x300
ocrRead the text in an image (tesseract)
padAdd a border or canvas: --border 20, --size 1080x1080
resizeScale: --size 800x, 1920x1080, 50%
rotateTurn clockwise by any angle (EXIF is applied first)
sharpenUnsharp mask: --radius 2
stripRemove EXIF, GPS and text metadata
to-gifAnimate images into a GIF: --fps 2
to-pdfOne page per image
to-videoSlideshow: --duration, --fps, --audio
transparentMake a colour see-through: --color white --fuzz 10
trimCut away plain borders, matching the corner colour
watermarkAdd text or a logo: --text, --image, --position

video

Trim, crop, speed, convert, compress, GIF, frames (ffmpeg)

infoDuration, size, frame rate, codecs
add-audioReplace or mix the soundtrack: --audio song.mp3
add-subtitlesAdd subtitles: burned into the picture, or --soft
chunkCut into equal parts: --length 10:00
compressSmaller file: --level, --target-size, --max-height
convertChange container and codec: --to mp4|webm|mov|mkv
cropCut a region: --box or --aspect 1:1
downloadSave a video: --resolution 720p, --to mp4|webm|mkv
extract-audioSave the soundtrack: --to mp3|m4a|wav|flac|ogg
extract-subtitlesSave an embedded subtitle track: --to srt|vtt
fadeFade: --fade-in 1 --fade-out 2
framesSave frames as images: --fps 1 or --at 0:05,0:10
from-imagesSlideshow: --duration, --fps, --audio
mergeJoin videos; --transition fades
muteRemove the audio track
padFit into a frame with bars: --size 1080x1920
repeatRepeat a video: --count 3 plays it three times
resizeScale: --size 1280x, x720, 50%
reversePlay backwards, picture and sound
rotateTurn clockwise: --angle 90
speedSpeed up or slow down: --factor 2 or --factor 0.5
splitCut out a part: --from 0:10 --to-time 0:25
stabilizeSteady shaky footage, or --denoise grain
stackPlay side by side: --direction
stripRemove metadata; no re-encode
thumbnailOne frame as an image, or --at a time
to-gifAnimated GIF: --fps, --size, --from, --to-time
watermarkAdd text or a logo: --text, --image, --position

audio

Convert, trim, merge, fade, speed, tag, normalize (ffmpeg)

infoDuration, codec, sample rate, channels
chunkCut into equal parts: --length 10:00
compressRe-encode as MP3 at --bitrate (kbps)
convertChange format: --to mp3|m4a|wav|flac|ogg|opus
denoiseReduce steady background noise such as hiss or hum
fadeFade in or out: --fade-in 2 --fade-out 3
from-videoSave the soundtrack: --to mp3|m4a|wav|flac|ogg
mergeJoin audio files in order
mixPlay files on top of each other, such as voice over music
normalizeEven out loudness (EBU R128)
remove-silenceCut silent parts: --threshold -40 --min-silence 0.5
reversePlay backwards
speedChange tempo, keeping pitch: --factor 1.5
splitCut out a part: --from 0:10 --to-time 0:25
stripRemove tags and cover art; no re-encode
tagSet tags: --title --artist --album --year, cover --image
volumeRaise or lower the level: --gain 6 or --gain -3 (dB)
waveformDraw the sound wave as an image: --size 1200x240

doc

Word/ODT/RTF: to text, Markdown, HTML, PDF; find and replace

infoTitle, author, words, headings, tables, images
from-htmlConvert the content to a Word document
from-mdWrite a Word document with headings, lists, tables, images
from-pdfEditable Word file, layout not kept
replaceFind and replace text in Word, keeping the formatting
to-htmlConvert to a web page
to-mdConvert to Markdown, keeping headings, lists and tables
to-pdfTypeset as PDF, content not layout
to-textExtract plain text

subtitle

SRT/VTT: convert, shift, burn in, extract

infoNumber of lines, start and end time
burnAdd subtitles: burned into the picture, or --soft
convertChange format: --to srt|vtt|txt|md
from-videoSave an embedded subtitle track: --to srt|vtt
shiftMove all timing: --by 2.5 (later) or --by -1 (earlier)

slides

PowerPoint: extract text, Markdown outline, PDF handout

infoSlide count, titles, images, notes
from-mdMake a deck: each heading or --- starts a slide
to-mdOutline in Markdown: a heading per slide, bullets, images
to-pdfHandout PDF of slide content
to-textExtract the text of every slide, form feed between slides

data

JSON/YAML/TOML/XML: convert, query with jq syntax, merge, diff

convertChange format: --to json|yaml|toml|xml|jsonl
diffCompare two documents by structure, across formats
formatPretty-print or minify: --compact, --sort-keys
mergeDeep-merge documents; later files win
queryQuery or edit with jq syntax: --expr '.items[].name'
validateCheck syntax, or a JSON Schema with --schema

sheet

Excel/CSV/ODS tables: filter, sort, join, pivot, convert

infoSheet names, rows, columns
convertChange format: --to csv|xlsx|json
dedupeDrop repeated rows, by whole row or --columns
extractTake a sheet or range: --sheet, --range A1:D20
filterKeep matching rows: --where "age > 30 and city = 'Oslo'"
from-pdfPull tables out by text position
groupGroup and total: --columns city --values sales --agg sum
joinJoin two tables on a key: --on id --how left
mergeStack tables; a shared header is kept once
pivotPivot table: --rows city --columns year --values sales
queryRun SQL over tables or a SQLite file: --sql 'select ...'
selectKeep, reorder or --drop columns
sortSort rows: --columns age,name --descending
splitSplit up: --sheets, --columns city or --every 1000
statsPer-column counts, empties, min, max, mean, sum
to-mdMarkdown table (first row is the header)
transposeSwap rows and columns

md

Markdown to PDF, HTML, Word, EPUB; table of contents

from-docxConvert to Markdown, keeping headings, lists and tables
from-htmlConvert to Markdown; scripts and styles are dropped
from-pdfText as Markdown paragraphs
to-docxWrite a Word document with headings, lists, tables, images
to-epubMake an e-book; each top-level heading starts a chapter
to-htmlRender to a styled HTML page, or --fragment
to-pdfTypeset Markdown as PDF
to-pptxMake a deck: each heading or --- starts a slide
tocInsert a table of contents from the headings

html

HTML to Markdown, text, Word, PDF; screenshots

screenshotRender a local page to an image with Chrome: --size 1280x800
to-docxConvert the content to a Word document
to-mdConvert to Markdown; scripts and styles are dropped
to-pdfMake a PDF; --browser keeps the CSS
to-textExtract the readable text

epub

E-books: info, to text, Markdown and PDF

infoTitle, author, language, chapters, images, words
from-mdMake an e-book; each top-level heading starts a chapter
to-mdConvert the whole book to Markdown
to-pdfTypeset the book as PDF
to-textExtract plain text in reading order

email

EML files: headers, text, attachments

infoFrom, to, subject, date, attachments
attachmentsSave the attached files
to-textThe message as text, with its headers

youtube

Info, download, audio, subtitles, thumbnail (yt-dlp)

infoTitle, duration, channel, chapters, qualities
audioSave only the sound: --to mp3|m4a|opus
downloadSave a video: --resolution 720p, --to mp4|webm|mkv
playlistList the videos of a playlist or channel
subtitlesTranscript: --lang en,zh --to md|txt|srt|vtt
thumbnailSave the cover image

file

Any file: hash, duplicates, batch rename, base64, encoding

base64Encode to base64, or --decode back to bytes
duplicatesFind files with identical content; deletes nothing
encodingConvert text encoding, normally to UTF-8; fixes line endings
hashChecksums: --algorithm md5|sha1|sha256|sha512|blake3|crc32
joinPut parts back together, in the order given
organizeCopy files into folders: --group type|ext|month
renameBatch rename: --find, --replace, --name, --case; --preview
splitCut a big file into parts: --part-size 100MB

qr

Create QR codes and barcodes; read QR codes

createMake a QR code or barcode: --text https://example.com
readDecode every QR code in an image

Any file

Commands outside a group

convertAny-to-any by extension, e.g. deft convert a.png -o a.pdf
infoType, size, pages, dimensions, duration
findSearch commands by intent
schemaJSON Schema of a command's inputs and options
engineManage external programs such as ffmpeg and yt-dlp
skillPrint or install the agent skill
doctorCheck the environment and engines

Built for the agent reading it.

Help in layers

An agent reads only what it needs: groups, then commands, then options. Each layer has a token budget that tests enforce.

$ deft find "burn subtitles"
subtitle burn  video -> video  Add subtitles: burned into the picture, or --soft

$ deft pdf
deft pdf - commands (input -> output)
  info         pdf -> report      Pages, size, metadata, fonts, images
  compress     pdf -> pdf         Make a PDF smaller: --level, --target-size 2MB
  delete       pdf -> pdf         Remove pages: --pages 2,5-7
  from-images  image... -> pdf    One page per image, in the given order
  from-md      md -> pdf          Typeset a PDF with headings, tables, code, images
  merge        pdf... -> pdf      Join PDFs in the given order
  number       pdf -> pdf         Add page numbers: --format "{n} of {total}"
  rotate       pdf -> pdf         Turn pages clockwise: --angle 90, --pages
  select       pdf -> pdf         Keep and reorder pages: --pages 3,1-2
  split        pdf -> pdf...      One file per page, or --every N, --ranges 1-3,4-
  to-images    pdf -> png|jpg...  Render pages to images: --dpi, --pages
  to-text      pdf -> txt         Extract text (form feed between pages)
  watermark    pdf -> pdf         Stamp text or an image on pages: --text, --image
More: deft pdf <command> --help

One line back

A result is one line that starts with ok, partial or error. Add --json for structure.

$ deft pdf compress photos.pdf
ok pdf.compress photos.pdf -> photos.compressed.pdf  704KB -> 272KB (-61%)  0.10s

$ deft pdf compress photos.pdf -o small.pdf --json
{"v":1,"ok":true,"op":"pdf.compress","items":[{"ok":true,"inputs":["photos.pdf"],"outputs":[{"path":"small.pdf","format":"pdf","bytes":278842}],"facts":{"bytes_before":720475,"bytes_after":278842,"ratio":0.387,"images_recompressed":1},"warnings":[]}],"elapsed_ms":67}

Errors that carry the fix

Every error has a stable code, and a hint when there is a concrete next step. The agent corrects itself in one turn.

$ deft pdf compress locked.pdf
error INPUT_ENCRYPTED locked.pdf: the PDF is password-protected
hint: add --password <password>

$ deft pdf compres photos.pdf
error UNKNOWN_COMMAND deft pdf has no command "compres"
hint: did you mean: deft pdf compress | deft audio compress | deft image compress

Local and careful

Inputs are never overwritten and writes are atomic. Deft collects no usage data. Only commands marked [net] touch the network.

$ deft pdf compress photos.pdf
error OUTPUT_EXISTS photos.compressed.pdf already exists
hint: pass --force to overwrite, or choose another -o

$ deft pdf compress photos.pdf --dry-run -o copy.pdf
plan pdf.compress photos.pdf -> copy.pdf

$ deft youtube info https://youtu.be/xxxx --offline
error OFFLINE deft youtube info needs the network, but --offline is set

The launch film.

Video slot: the launch film goes here. Add site/media/launch.mp4 before hosting.

Build it today.

No release exists yet, so today you build from source. Video and audio commands need ffmpeg. Install guide.

git clone https://github.com/firstprayer/deft
cd deft && cargo install --path crates/deft-cli --locked
deft doctor
deft skill install --agent allTeaches Claude Code and Codex that deft exists.

Release binariesPlanned
Install scriptPlanned
Homebrew tapPlanned
cargo install deftcliPlanned