Frequently asked questions

Perplexity routes across models too — but its routing serves their margin: it's opaque, resets per thread, and is documented to silently downgrade your model under load. Tokenokio routes on YOUR accounts (your Claude/ChatGPT plans and your keys), shows the model and the reason for every call, reports the real cost and what the top model would have charged, escalates UP when an answer is weak instead of down under load, and lets you export the whole history. Same idea, opposite incentive.

No. Everything core runs in your browser: scanning, reports, optimisation, and live routing with an API key — no install, no signup, and your usage data stays on your device. One optional extra: running prompts on your Claude/ChatGPT plans at $0 uses the Tokenokio desktop app — one download, no terminal (Mac first, in final signing). A terminal path exists for advanced users who prefer it; you'll never need it.

Yes — that's the exact problem it was built around. Scan your usage (60 seconds, local, free) to see which models and projects consumed your quota, then let the router send small jobs to cheap models on your own keys so plan quota is saved for the heavy work. The full playbook is at tokenok.io/usage-limits.

Independent AI spend intelligence. Tokenokio shows you every AI token you use — across Claude, OpenAI, Grok, DeepSeek and more — priced in dollars, attributed by project, with the recoverable waste called out. Scan your machine in 60 seconds; nothing is uploaded.

Three ways, all under a minute: (1) Scan — pick your ~/.claude/projects or ~/.codex/sessions folder on the Track page; it parses in your browser. (2) Sync — paste a read-only reporting key (OpenRouter, OpenAI Admin, Anthropic Admin). (3) Import — drop a CSV or JSON from any provider.

Press Cmd+Shift+. in the folder picker to reveal hidden folders, then choose ~/.claude/projects (Claude Code) or ~/.codex/sessions (Codex). Safari and Firefox can't scan folders yet — run the one-line terminal scan shown on the Track page instead, then import the JSON it creates.

Routing: the full September 2026 frontier — Claude Fable 5.1/Opus 5/Sonnet 5, GPT-6 Astra and the GPT-5.6 family, Gemini 3.8, DeepSeek V4.1, Kimi K3, Grok 4.6, Qwen 3.8, GLM 5.3, Meta Muse Spark, and TypeSafe's Jev decision model — plus anything new the Model Radar spots on the live market feed, routable without waiting for an update. Local scans: Claude Code and OpenAI Codex today. Key sync: OpenRouter, OpenAI, and Anthropic. Anything else arrives via CSV import.

Jev is the first System One decision model: instead of generating text, it returns typed judgments — a choice from your options, a score, or a yes/no probability — with calibrated confidence, in a few hundred milliseconds, at $0.042 per million input tokens (output is free). A huge share of real-world AI calls are exactly that shape: route this ticket, is this spam, which option is right. Tokenokio's router sends only that kind of work to Jev — never writing, code or reasoning, which it can't do — so classification jobs that cost cents on an LLM cost fractions of a cent, with the confidence score feeding our escalation ladder: low confidence and the job re-runs on a stronger model automatically.

Two grades, always labelled. Provider-reported: real dollars from the provider's own billing API. Estimated: token counts priced at public list rates — shown when you're on a subscription plan or the provider doesn't report cost. Every number tells you which grade it is.

On your device. Scans parse entirely in your browser and the results are stored in your browser's local storage. Export it or clear it any time from the Track page. Tokenokio's servers keep nothing.

Never. Scans read token counts, models, dates and project names — the accounting lines, not the content. Prompts, responses and files never leave your machine.

Reporting keys are read-only — they can view usage, never run models or move money. They're stored in your browser, sent once per sync over HTTPS, used in memory, and never stored or logged by us. Remove them any time.

Security is designed in, not bolted on: we're local-first and least-trust, so there's almost nothing to attack. Your keys, tokens, prompts and payments never touch our servers, and there's no user database to breach — access is stateless and cryptographically signed. Every surface has been through an adversarial security review modelled on AI-assisted and state-actor-level attackers — the local meter, the access and licensing cryptography, billing, and the web app — with findings remediated and re-verified, and security behaviour covered by automated tests that run on every change. Controls include fail-closed access, payment-bound licensing, an HTTPS-only proxy, strict output escaping and consent-gated analytics that carry no keys or content. Summary and our responsible-disclosure policy: SECURITY.md and security@tokenok.io.

Minimal by architecture, not by promise: your usage data, prompts, keys and payments never touch our servers, and there is no user database to breach — access is stateless and cryptographically signed. What an attacker could take from us is close to nothing of yours. Every surface has been through adversarial security review with findings fixed and re-verified, and security behaviour is enforced by automated tests on every change. Details and our responsible-disclosure policy: SECURITY.md and security@tokenok.io.

Yes — by architecture. We minimise to near zero: your usage data stays on your device, analytics runs only with your consent, and we honour Global Privacy Control. Full detail, including your rights in the EU, UK and every US state, is in our Privacy Policy.

You stay in full control. Your provider accounts, keys and AI content remain yours under your agreements with those providers. Tokenokio reads usage numbers you choose to share with your own browser — see the Terms for the plain-English version of this.

The scan, menubar/CLI and core reports are free — forever. Pro is US$15 one-time (waste deep-dive, budgets, exports, project attribution — no subscription). Team is US$99/month (org sync, team showback). Team+ from $199/month, enterprise audits from $2k. The goal is simple: save you far more than you pay.

Enter it on the Pricing page (or when the app asks). It unlocks full access instantly — no card required during beta.

Card via Stripe, monthly, cancel anytime. Beta users get generous notice and founding-member pricing before anything changes. You'll never be charged silently.

Subscriptions for the intelligence, plus enterprise audits. We deliberately never resell tokens, hold credits or handle your provider payments — zero float means our only incentive is saving you money.

The near roadmap: live budget alerts and agent kill-switches, realised-savings tracking (measured, not estimated), team showback, and deeper attribution. Paid plans fund the build and get everything first.

Your recoverable spend, categorised: cache misses, verbose outputs, frontier models on simple work, and runaway-agent spikes — each priced in $/month. Claims are deliberately conservative, capped, and show their assumptions.

The same task, priced on every model — in tokens and dollars. The spread is routinely 100× or more. It's how you know a cheaper route exists before you commit to it.

Yes. The Router decomposes a complex prompt — research, analysis, writing, code — and routes each part to its best-value model, priced against running everything on one frontier model.

Set a monthly figure on the Track page. Tokenokio projects your month-end spend from real usage and turns amber at 80%, red when you're on course to breach — with the overage in dollars.

A finance-ready export: day, attribution (project/workspace/key), provider, model, tokens, cost, and whether each cost is provider-reported or estimated. Built for the person who approves the budget.

Not to see your own results — scan and go. Accounts arrive with paid plans so your settings follow you across devices; your usage data can stay local even then.

Yes. Scan each machine, export the JSON, and import them into one Track page — results merge. One unified ledger across every machine and provider.

The three fixes that solve most things: refresh the page, re-run the scan (Cmd+Shift+. reveals hidden folders), and check the key you pasted is a reporting/admin key rather than a standard inference key. Still stuck? Escalate below and a human will help.

Track page → Clear data. That wipes local storage instantly. Remove any pasted keys in the same panel. Because we store nothing server-side, that's genuinely everything.