Equalang

Point it at a file and it uploads the file, waits for the job, and writes the results beside the source. You get paths back, never the file's contents - layout, tables, images and formulas survive, and nothing large enters the conversation.

Work spends the user's credits. Say what it will cost and get their agreement before translate or transcribe. estimate gives the number for a file, for free.

Setup

The key is one line, EQUALANG_API_KEY=el_..., in ~/.config/equalang/.env (under $XDG_CONFIG_HOME if that is set): one file per machine, shared with the Equalang MCP server and kept when this skill is reinstalled or updated. The user creates the key at https://equalang.com/api-keys.

dir="${XDG_CONFIG_HOME:-$HOME/.config}/equalang"
mkdir -p "$dir" && (umask 077 && echo 'EQUALANG_API_KEY=el_...' > "$dir/.env")

EQUALANG_API_KEY in the environment wins over the file. Without a key every command answers with how to get one and where it goes. Never invent a key - ask the user.

Commands

Run from this skill's directory; python3 and nothing else is needed. Every command prints one JSON object; --help lists a command's flags.

# What would it cost? Uploads without starting anything; free. `kept_until` says how long the upload is held.
python3 scripts/equalang.py estimate report.pdf
#   -> {"file_id": "...", "credits_to_translate": 12.34, "credits_to_transcribe": null, ...}

# Translate a file. A path, a public URL (Equalang fetches it), or --file-id from estimate.
python3 scripts/equalang.py translate report.pdf --to zh-CN
python3 scripts/equalang.py translate --file-id <file_id> --to zh-CN        # no second upload
python3 scripts/equalang.py translate https://example.com/deck.pptx --to ja -o ./out

# A recording: translated subtitles, or what was said in the language spoken.
python3 scripts/equalang.py translate talk.mp4 --to en --bilingual --format srt --format txt
python3 scripts/equalang.py transcribe interview.mp3 --format srt --format txt

# Strings in bulk (each translated on its own; at most 50 of 5,000 characters, 20,000 per call).
python3 scripts/equalang.py text "Save changes" "Delete project" --to de
# One long plain text - an article, notes, a .md - up to 100,000 characters. Equalang cuts it at sentences: never split it yourself.
python3 scripts/equalang.py text --file notes.md --to ja -o notes.ja.md

# Look a job up again · stop one · the balance · language codes (no key needed)
python3 scripts/equalang.py status <job_id> -o ./out
python3 scripts/equalang.py cancel <job_id>
python3 scripts/equalang.py balance
python3 scripts/equalang.py languages chinese

Reading the output

A finished job prints outputs (each with its path) and credits_charged. Tell the user both. Do not read a produced file back to summarise it unless asked: it is a whole document, and the point of a path is that its contents stay out of the conversation.

A failure prints {"error", "code", "retryable"} and exits 1. Retry only when retryable is true; INSUFFICIENT_CREDITS means the user tops up at https://equalang.com/pricing, not that you try again.

Worth knowing

Formats. Documents: pdf, docx, pptx, xlsx, epub, html, txt. Subtitles: srt, vtt. Pictures: jpg, png, webp, bmp. Recordings: mp3, m4a, wav, flac, ogg, aac, opus, mp4, mov, webm, mkv. Up to 100 MB. transcribe takes recordings only.
Results never overwrite. They land beside a local source (or in -o DIR, or the current directory for a URL) under the name Equalang gives them; a taken name gets (1). Asking status again about a finished job names the files it already wrote rather than copying them.
Minutes, not seconds, for long files. The command waits and prints progress to stderr. Interrupting it does not cancel the job: status <job_id> picks it up, cancel <job_id> stops it. A cancelled or failed job is not charged.
A recording is charged for the speech actually heard - silence and music are not - so it usually costs less than estimate said; a document costs what estimate said.
Language codes look like en, zh-CN, zh-TW, ja, pt. They are read from the live API, never from memory: languages lists the ones files take, languages --kind text the wider set text takes, and languages chinese finds one by name. Omit --source-language to have it detected.
Several files? Estimate each, add the numbers up, and ask once - never loop over a folder on your own say-so.
The API behind this, for anything the script does not do: https://equalang.com/llms.txt