CloTokO — LLM Token Counter & Prompt Cost Optimizer

CloTokO measures your prompt and cuts what you should not pay for — live, as you type, and 100% offline: no network calls, no telemetry, no dependencies.

In Copilot Chat: @clotoko

Type @clotoko followed by a prompt and get the whole picture before you spend anything on it:

ISSUES · 272 tokens · $0.00040/call · $0.6012/mo

| | now | optimised | saving |
|---|---|---|---|
| tokens  | 1,427 | 255  | -82%     |
| cost/mo | $0.86 | $0.15| -$0.71   |
| CO2e/mo | 12 g  | 2 g  | -10 g    |

Energy: about 8x boiling a kettle, or driving 2.1 km every month (estimated).

X ~58% chance this needs a second attempt — state the exact output shape
X 90% of this call is the answer, not your prompt — cap the answer length
! 4 different jobs in one prompt — split them into separate calls

Prefer chat only, with no editor decorations? Set clotoko.surface to "chat".

The analysis, not just a word count

Trimming filler is the small half of the problem. CloTokO also reports the things that actually decide what a prompt costs and whether it works:

• Where the money really goes. On a typical call the answer is 90%+ of the bill, so deleting words from the prompt moves almost nothing. It says so, with your numbers.
• Retry risk. Vague scope, no output format and no length limit make a second attempt likely — and a retry doubles the call. It quotes the phrases responsible.
• Contradictions, e.g. asking for a brief answer and an exhaustive one at once.
• Too many asks in a single call, which reliably produces partial answers.
• Prompt-cache ordering — providers discount cached input by 50-90%, but only match a prefix, so variable content near the top forfeits the discount.
• Energy and carbon, scaled by model tier, with a comparison you can picture. Set clotoko.gridIntensity for your region (default 400 g CO2e/kWh).

It will not silently damage your prompt

Every rewrite is checked before you are offered it. Numbers and product names are load-bearing: if 3 scenarios, ESUs or PostgreSQL would be lost, the rewrite is discarded and CloTokO tells you why rather than reporting a saving that never existed.

What it does while you work

You write a prompt. CloTokO continuously shows what it costs and what to cut.

User: Please could you kindly, in order to help me, review this.   142 tok · $0.0021 · −38% possible

It tells you which model to use

CloTokO classifies the task (11 categories: reasoning, architecture, code-gen, analysis, QA, summarisation, rewriting, translation, extraction, classification, creative), estimates the capability it actually demands, then scores every model on capability fit, context headroom and price. It recommends the cheapest model that clears the bar — and shows its reasoning.

Task looks like classification / tagging ("sentiment").
That needs roughly 31/100 capability - raised by strict requirements.
Cheapest model clearing that bar: gpt-5-nano (50/100, $3.67/mo).
You are currently paying 190.9x more than necessary.

Habitual boilerplate is stripped before classifying. "Think step by step", "take a deep breath" and "you are a world-class expert" are instructions to the model, not evidence the work is hard — counting them would recommend a frontier model for sentiment tagging.

It writes you a better prompt

Beyond deleting waste, CloTokO reconstructs the prompt in a compact canonical shape:

It never returns a longer prompt — on an already-lean prompt it says so and leaves it alone. Your data is never paraphrased; only instruction prose is rewritten.

Press Ctrl+Alt+W to apply, or preview the diff first. It always confirms before replacing.

It knows where your prompt starts and ends

CloTokO never costs "the whole file". It finds the prompt block under your cursor:

• Chat transcripts — splits on System: / User: / Assistant: markers, so each turn is priced separately.
• Markdown — falls back to --- rules, then headings.
• Python / JS / TS — finds prompt string literals inside your code and scores them, so a file path or a const SECRET = "ABC..." is never mistaken for a prompt.

Optimize then rewrites only that prompt — it will not touch your system prompt or the surrounding code, and in Python/JS it edits the string contents without disturbing the quotes.

Why prompts cost more than they should

CloTokO measures all six on the prompt in front of you and fixes what it can automatically.

Waste rules

~25 rules with one-click fixes, in three safety tiers:

Strategy recommendations

Ranked by impact, each with the concrete action and expected saving:

• Output caps — output costs 4–8× input on every major model
• Prompt caching — reorder so the static prefix is byte-identical; shows the exact per-call saving for your model
• Batch API — 50% off anything non-interactive
• Model right-sizing — scores task complexity, names the cheaper model and the cascade pattern
• RAG instead of stuffing — 80–95% input reduction
• History windowing — flattens O(n²) chat cost to O(n)
• Few-shot trimming — accuracy saturates at 2–4 examples
• Semantic caching, streaming + stop sequences, structured output to kill retries
• Token density warnings — id/hash-heavy content, non-Latin script (2–3× cost)
• Context-window overflow risk

Token budgets

Set CloTokO.tokenBudget (or run CloTokO: Set Token Budget). Any prompt over it gets an orange inline marker, a status-bar flame, a CodeLens, and a Problems entry telling you the monthly cost of the overshoot.

Commands

Settings

Accuracy

Token counts come from an offline model of the GPT pre-tokenizer plus BPE merge heuristics and a frequent-word table, calibrated per encoding (o200k, cl100k, Claude, Gemini, Llama). On English prose, markdown, JSON and code it lands within about ±5% of tiktoken — with no 2 MB vocabulary download and no network.

Prices are planning defaults (see PRICES_AS_OF in src/models.js). Override them in CloTokO.customPrices.

Privacy

No network access, no telemetry, no dependencies. Your prompts never leave the machine. The only persisted state is a local savings counter in VS Code's globalState.

Install

Copy the folder to %USERPROFILE%\.vscode\extensions\ and reload VS Code, or install the included CloTokO-1.1.0.vsix. No build step — plain CommonJS JavaScript.

Bonus: terminal CLI

$env:ELECTRON_RUN_AS_NODE=1
& "$env:LOCALAPPDATA\Programs\Microsoft VS Code\Code.exe" tools\cli.js demo\wasteful-prompt.md gpt-4o
# flags: --write (emit file.optimized.md)  --template (print the prompt skeleton)

Try it

Open demo\chat-transcript.md — each turn is priced separately. Put your cursor in a turn and press Ctrl+Alt+I, then edit and watch it update. demo\prompts.py shows detection inside Python string literals.

Tests

$env:ELECTRON_RUN_AS_NODE=1
& "$env:LOCALAPPDATA\Programs\Microsoft VS Code\Code.exe" test\run-tests.js
& "$env:LOCALAPPDATA\Programs\Microsoft VS Code\Code.exe" test\activate-test.js

135 assertions: 99 unit (tokenizer, pricing, every rule, prompt scoping, offset mapping, optimizer, model advisor, prompt rewriter, caching, regression guards) and 36 activation tests that load extension.js against a stubbed VS Code API and verify the live decorations, cursor scoping, budget warnings and every command.

One test asserts no source file imports http, net, child_process or calls fetch — the offline guarantee is enforced, not just claimed.

License

MIT