102 free tools · nothing you type is uploaded
tokens — that sentence, as the model actually reads it. Every colour is one token you pay for.
Everything an LLM charges you for is measured in tokens, and tokens are the one unit you cannot see in your editor. This site makes that unit visible — then lets you build prompts, agents and skills that show what they cost before you ship them.
How it works
There is nothing to install, configure or sign into. Open a tool, paste what you have, read the number.
102 of them, grouped by what they do. Press ⌘K from any page to search by name, or browse the tools index by category.
Not a sample — the actual prompt, payload or document. It never leaves your browser, so there is no reason to sanitise it first. That is the whole point of running client-side.
Exact where the tokenizer is public, clearly labelled as an estimate where it is not. Every result links onward to the tool that answers whatever it made you ask next.
Try it here
The same encoder the API bills you with, running on your machine. Change a word and the count moves — including on models where the honest answer is that we are estimating.
Open the full counter →20tokens
What you can do here
Twelve things people actually arrive wanting to do. Each one goes straight to the tool that does it.
Why any of this matters
Each of these is a tool on the site. Here is what each one is actually about.
01
Text is converted into entries from a fixed vocabulary before the model sees it. Common English words are one token; a UUID is a dozen; indentation is one per level per line.
English prose runs about four characters per token. JSON runs 2.2, and non-Latin scripts can fall below 1.6 — which is why estimating from character count misleads you most on the payloads that dominate your bill.
See your own text split →02
System prompt, tool definitions, conversation history, retrieved documents and the answer the model is about to write all compete for the same space. Not two budgets — one.
Fill it with input and there is no room left to answer. Fill it at all and recall starts degrading, which is why a tightly retrieved 20,000 token prompt regularly beats a padded 500,000 token one.
Check what fits →03
The API is stateless, so every turn resends the whole conversation. Turn ten pays for turns one through nine again.
Total input therefore grows with the square of the turn count. A per-request estimate multiplied by turns is commonly wrong by three to four times, in the expensive direction.
Model your own conversation →04
System prompt plus every tool name, description and schema is sent on every iteration, whether or not a tool is called. On a dozen tools that routinely exceeds a thousand tokens per request.
Multiply that by an iteration limit and add the accumulated history, and you have the number almost no framework shows you.
Simulate the loop →05
Eight chunks of three hundred tokens is 2,400 tokens on every request, billed every time — and the model typically cites two or three of them.
Retrieving fewer, better chunks cuts the largest line of a RAG bill and usually improves the answer, because irrelevant context degrades output as well as costing money.
Price your retrieval →Build
Not another article about prompting. Actual builders that produce a file you can ship, and tell you the per-request cost while you write it.
Six sections in the order models weight most heavily. XML delimiters, token count and cost as you type.
<instructions>
Classify the ticket into
billing, technical or account.
</instructions>
<output_format>
{"category": string}
</output_format>Model, system prompt, tools and iteration limits — with the overhead your config adds to every single request.
{
"name": "support-triage",
"model": "claude-sonnet-5",
"max_iterations": 6,
"tools": ["search_docs"]
}Validated frontmatter, trigger quality checks, and the split between what is always in context and what loads on demand.
--- name: token-cost-review description: Use when the user asks to reduce token usage ---
Who it is for
Measure a real request in the token counter, then project it in the cost calculator. If your product is conversational, the chat cost estimator will explain the gap — per-request maths understates chat by three to four times.
Filter by budget and context in the model finder, check every price against its source in the price table, then read how to read the numbers — the headline rate decides a minority of real bills.
Compose in the agent builder and the skill builder, which show the token overhead your configuration adds to every request — the number almost no framework surfaces.
Start with token, then context window. The glossary has 84 terms, each with the mistake people usually make with it.
102 tools
Learn
The tools give you a number. These explain what to do about it — whether you need a procedure, a definition, working code or a list to tick off before something ships.
6
Procedures. Numbered steps, the mistakes that make each one fail, and a tool at the point you would otherwise go looking for one.
160
8 lists of things you can verify rather than agree with, each item with the reason it is there. Ticks persist in your browser, so a launch list survives closing the tab.
10
Python and TypeScript for the parts everyone writes: counting tokens, streaming, retry with backoff, tool loops. Including the error handling the documentation snippet leaves out.
84
Definitions that say what the thing costs and what it trades off — plus the mistake almost everyone makes with it, which is usually what you came for.
9
Longer essays on why something is true — how tokenization works, why chat costs compound, what caching actually saves.
42
Input, cached input and output rates for every model tracked, each with the provider URL it was read from and the date it was read.
Keep your work
Anything you make in the prompt, agent or skill builders can be saved to your library with one click. It lives in this browser — no sign-up, no upload, and nobody including us can read it.
The trade is honest: a library stored locally does not follow you to another device. That is what the export button is for, and the file it produces is plain JSON that you own.
Open your library →The difference
The genuine o200k_base and cl100k_base encodings run on your machine. For 18 of the 42 models here the count is exact — the same number the API bills.
Anthropic, Google and DeepSeek publish no browser tokenizer. Rather than quietly substituting OpenAI’s, every one of those figures carries a badge and a written derivation on the methodology page.
Each entry stores the provider URL it came from and the day it was read, published in the price table. Entries over 45 days old are flagged. Stale pricing is worse than none.
Static files, no application server, no account. Load a page, disconnect from the network, and every tool keeps working — a demonstration rather than a promise.
Questions
Free, no sign-up, and nothing you type is transmitted — because there is no server here to transmit it to.