Tokly: AI Token Optimizer

Pasting files into an AI chat? Tokly makes them cost fewer tokens first — on your phone, offline, and without your files ever leaving the device.

Pick a file and Tokly shows its real token count. It then tries many representations — compact JSON, TOON, compact tables, clean Markdown, compressed logs — counts each one with a real tokenizer, checks it against your file and keeps the smallest one that passes. A live processing screen shows every step and the best result so far.

THREE MODES
• Lossless (default): your content stays intact. An output either keeps your text as it is, or is decoded back and compared with your file — same values, same cells, same visible text — before it can win. It may use light Tokly notation, such as a one-value column written once or ×N for repeated log lines, explained inside the output. Web pages become clean Markdown without scripts and styles, with every visible word kept; Maximum can also drop navigation, headers and footers. In PDFs, page headers and footers repeated on most pages are removed.
• Maximum: also allows denser Tokly notation, such as a mark for “same value as the row above” or record tables for mixed JSON, with a short explanation inside the output. It also tries options that drop details like code comments or links. These results are always clearly labeled.
• Context: keeps only what one purpose needs — a code map (signatures and types), an outline, tables only, or text without repeats. Incomplete on purpose, with a count of what was left out (for example, 19 function bodies).

TOON AND COMPACT TABLES
• Uniform JSON arrays can become TOON, an open format made for LLMs: one header, then one row per object — used whenever it is smaller
• A column with the same value in every row is written only once
• In logs, runs of identical lines become ×N and a shared line prefix is written once

HONEST NUMBERS
• Exact counts with OpenAI’s open o200k and cl100k tokenizers, running locally; a clearly labeled estimate for Claude, Gemini and other models
• Net savings: explanation lines are part of the output and are counted, and Tokly shows how many tokens they take
• On our 20-file test set, a pretty-printed JSON API response holding 200 uniform user records went from 35,200 to 17,101 o200k tokens in Lossless mode: 51% fewer, and 32% fewer than compact JSON. Your results depend on the file.
• If a file is already efficient, Tokly says so instead of inventing savings

WHAT IT HANDLES
• PDF (text layer), Word (DOCX), Excel (XLSX), PowerPoint (PPTX)
• TXT, logs, Markdown, JSON, JSONL, CSV, TSV, HTML, XML, YAML
• Source code in 20+ languages

FREE AND PRIVATE
• Free, with no account, no ads, no analytics and no tracking
• Works offline: everything is processed on your device and nothing is uploaded
• Optional local history (off by default) keeps only a summary — file name, type, token counts and date — never contents

Copy the result straight into your AI chat, optionally tagged with the file name so the model knows what it is reading, or share it as a file — or pick any other validated format yourself.

Tokly is an independent app and is not affiliated with or endorsed by OpenAI, Anthropic or Google. OpenAI, Claude and Gemini are trademarks of their respective owners.