Privacy Policy

This policy applies to personal data we collect when you visit tokenzip.ai or use the TokenZip product (together, the “Service”). “Personal data” means information that identifies or could reasonably be linked to you.

It does not cover the practices of companies we don’t control — for example, Stripe’s checkout pages, Google’s sign-in screens, or the AI providers whose accounts you connect to TokenZip. Those services have their own privacy policies. For how we handle the usage data you connect and which of our providers hold it, see section 5; for the one feature that sends your content to an AI model, see section 6.

Profile and contact data

When you create an account we collect your email address and, if you sign in with Google, the name and profile details Google shares with us. Authentication is handled by Clerk; if you sign up with a password it is stored only in hashed form by Clerk — we never see or store it in plain text. If you send us a message through the contact form on our site, we store your name, email address, your message, and a hashed form of your IP address, which we keep only to rate-limit abuse of the form.

How long we keep contact-form messages. A contact-form submission exists in two places, and they are cleaned up differently, so it is worth being exact. The copy stored in our systems is deleted automatically two years after you send it; the rate-limit counters behind the form are deleted after 30 days. Your message is also delivered by email to our own inbox, and that delivered copy is not covered by the automatic deletion — it is removed by hand when you ask us to delete your message, within the same 30 days as any other deletion request. This is all stated separately because you do not need an account to use that form: if you never become a customer there is no account for you to close, so the retention in section 8 would never reach you. Ask us at hello@tokenzip.ai and we will delete both copies.

Payment data

Payments are processed by Stripe. Your card number is entered directly with Stripe and never touches our servers. We receive only limited billing metadata: your plan, subscription status, and Stripe’s identifiers for your customer record and subscription.

Provider credentials

If you connect an AI provider account — OpenAI or Anthropic — we store the API key you give us encrypted at rest with AES-256-GCM, and decrypt it only to read the usage records you asked us to audit. We never write to your provider account. You can revoke the key at your provider at any time, which immediately ends anything TokenZip can do with it; to have the stored credential itself removed, email us (see section 8 for how removal works and how long it takes).

Usage data you connect

We collect the AI usage records you connect or upload — from a provider API, or from an export produced by Helicone, Langfuse, LangSmith or a generic JSONL file. Those records contain call metadata such as model, timestamp, token counts and cost, and, depending on the source, the prompt and completion text your own systems produced. We also store the audit reports we generate from them. Where those records go, and which of our providers hold them, is set out in section 5.

Device and server logs

Our servers keep basic request logs — IP address, browser and device type, the path requested and when — which we use to operate, secure and debug the Service. We do not run a product-analytics tool on top of them (see section 4).

Where this data comes from

• Directly from you — when you create an account, connect a data source, upload a usage export, or contact us.
• Automatically — through our servers’ request logs as you use the Service.
• From service providers — Stripe sends us billing events; Clerk sends us your account profile; Google sends us basic profile information if you choose Google sign-in; a connected AI provider returns the usage records we request with your key.

• Providing the Service — creating your account, ingesting and analyzing the usage data you connect, generating and storing your audit reports, running replay where you have enabled it, and processing payments.
• Improving the Service — diagnosing failures, understanding which parts of the analysis pipeline are slow or wrong, and improving detection quality.
• Communicating with you — transactional email such as sign-in confirmations, billing receipts, notifications that a run has finished, and responses to your support requests.
• Security and fraud prevention — protecting the Service, our users, and others.
• Legal compliance — meeting our obligations under applicable law.

We do not sell your personal data, and we do not share it with third parties for their own advertising. We do not train AI models on the usage data you connect or the prompts inside it. Replay traffic is routed with a no-prompt-training constraint, and the providers it can reach are selected under commercial API terms that do not grant training rights — section 6 says exactly what that control is and what it does not cover. If we ever want to use your data for a materially different purpose, we’ll tell you first.

We use essential cookies only. They are set by Clerk, our authentication provider, to keep you signed in; the Service doesn’t work without them.

We collect anonymous, cookieless, aggregate page-view analytics through Vercel Web Analytics, our hosting subprocessor — no cookies, no cross-site tracking, and no individual profiles. We do not run a product-analytics tool that tracks individual users, we do not set advertising or cross-site tracking cookies, and we do not run website experiments that follow you between visits. If that changes, this section and the list in section 7 will be updated before it does. Because there is no consistent industry standard, we do not currently respond to “Do Not Track” browser signals. You can block or delete cookies in your browser settings; if you block essential cookies, signing in will stop working.

TokenZip’s core feature is analyzing your own AI usage. When you connect a provider API key, we read the usage records that key exposes. When you upload an export, we read the records in the file you gave us. We only ever read the sources you point us at.

Those records sometimes contain personal data, because the prompts and completions your own systems produced may contain it. Where such content appears, we process it solely to run your analysis and produce your report. We do not use it to build profiles of individuals, we do not train models on it, and we do not sell it. A report itself shows only an allowlisted subset of what we hold — dollar figures, token counts, abbreviated prompt signatures, and short replay excerpts — rather than a dump of your data.

Where it is held. Running the Service means your records pass through infrastructure providers, and we would rather name that plainly than let “we don’t send your data anywhere” imply otherwise:

• Uploaded exports. A file you upload goes from your browser straight to Vercel Blob, our object store. Objects there sit at long, unguessable URLs but are public-addressed — anyone holding the URL could fetch the object — so they are treated as transient. We delete the object as soon as the export has been ingested; if that delete fails, the object stays queued and is retried by a daily job until deletion is confirmed. An upload nobody ever runs is deleted by that same daily job once it is more than 24 hours old, which in the worst case means it is removed within 24 to 48 hours of upload rather than exactly 24. Moving prompt-bearing uploads to private storage is a planned change; until it ships, this is what happens.
• Parsed records and reports. Once ingested, your usage records — including any prompt and completion text they carry — are stored in our Postgres database at Neon, along with the reports built from them. They stay there until removed (see section 8).
• Everything else. Serving the pages and APIs you use runs on Vercel. Neither Vercel nor Neon is a model provider: nothing described here sends your content to an AI model. The only thing that does is replay, in section 6.

What you must not connect. Because a replayed prompt is forwarded to third-party model providers under their own terms, our Terms of Service prohibit connecting usage data that contains sensitive personal information — health information, financial account numbers, government identifiers and the like. If your traces may contain those categories, do not enable replay for that workspace.

As between you and us, the usage data you connect stays yours and we handle it on your instructions — we decide nothing about what is in it, and we process it only to run the audits you ask for. If you are an end user of a TokenZip customer’s application and believe a report contains your personal data, contact that customer, and you can also write to us at hello@tokenzip.ai — we’ll review promptly and pass the request on.

Replay is the only feature that sends your content to an AI model provider, so it gets its own section. It is off unless someone in your workspace turns it on, and with it off TokenZip makes no model inference call with your content at all. That is a narrower statement than “nothing leaves”: hosting and storage providers still hold your records either way, which is what section 5 describes.

• It is opt-in, per workspace. Replay is off until someone in your workspace turns it on. Until then no prompt, system instruction or completion text you connect is submitted to any model.
• What it does. When it is on, we re-execute a sample of your recorded calls on an alternative model, then ask a third model to score whether that alternative’s output could have replaced the original — so a saving on your report can be measured rather than estimated.
• Exactly what is sent. For each sampled call, four pieces of your content are transmitted: the task prompt, the system instructions that accompanied it, the original completion your own production model returned, and the candidate completion the alternative model produced from that prompt. The scoring step receives all four, because judging substitutability means comparing the two outputs. The model identifiers involved are sent with them. Your profile, billing and credential data are never sent.
• Where it goes, and the routing constraint. That content is sent through Vercel AI Gateway to the model providers behind it. Every request TokenZip makes sets the gateway’s no-prompt-training routing constraint, so it is directed to providers that do not train on prompt data, and those providers are reached under commercial API terms that do not grant training rights. We do not currently require zero-retention routing, which is a separate and stronger control: a provider may hold the request for its own abuse-monitoring and retention periods under its terms.
• How much of it. Only a sample — a capped number of calls per finding, under a token budget for the whole run — not your full dataset.
• We screen for a few specific patterns. Before a sample is sent, we check it for specific structured patterns — checksum-valid payment card numbers and certain government identifier formats among them. This is a narrow pattern match on particular formats, not a review of what your data contains.
• What a match does, at each of two stages. If the match is in your own content — the prompt, the system instructions or the original completion — the sample is dropped before anything is sent, so it never reaches a model at all. If instead the candidate model’s reply matches, the candidate call has already happened: we discard that sample before the scoring step rather than pass it on, but we cannot un-send the call that produced it. Either way the sample is excluded from the evidence and the finding reports fewer samples rather than quietly including it.
• It cannot see most of what matters. The check cannot detect sensitive information written as ordinary prose — a clinical note, someone’s diagnosis, a description of their finances — and it does not recognise every format of every identifier. It is not a classification system and it does not make replay safe for sensitive data. Do not enable replay if your traces may contain health, financial, or identity information about individuals. The rule in our Terms of Service stands unchanged, and only you can see what is actually in your traces.
• Turning it off. You can disable replay for your workspace at any time, which stops all future replays. Replay excerpts already stored on past reports stay until they are removed (see section 8).

We share personal data only with the service providers we need to run TokenZip, each bound by its own contractual and privacy obligations:

• Vercel — hosting and content delivery; Vercel Blob, the object store an uploaded export passes through (see section 5); the AI Gateway that routes replay requests (see section 6); and Vercel Web Analytics, the cookieless, aggregate page-view analytics described in section 4.
• Neon — the Postgres database holding your account, your usage records including any prompt and completion text in them, and your reports.
• Clerk — authentication and account management.
• Stripe — payment processing.
• Resend — delivery of transactional and contact-form email.
• AI model providers — only where your workspace has enabled replay, and only the four content fields enumerated in section 6. They do not receive your profile or payment data.
• Google — only if you choose to sign in with Google.

Beyond service providers, we may disclose personal data when required by law or legal process, to enforce our Terms of Service, or to protect the rights, safety, or property of TokenZip, our users, or others. If TokenZip is involved in a merger, acquisition, or sale of assets, personal data may transfer as part of that transaction — we’ll notify you of any resulting change in ownership or use of your personal data.

We may also create and share aggregated or de-identified data (for example, overall usage statistics or model-price benchmarks) that cannot reasonably be used to identify you or to reconstruct your prompts.

We use industry-standard safeguards appropriate to the data we handle: encryption in transit, AES-256-GCM encryption at rest for the provider API keys you give us, hashed credentials at our authentication provider, access controls, and infrastructure providers with strong security programs. No method of transmission or storage is completely secure, so we can’t guarantee absolute security — please use a strong, unique password.

Reports are published at long, unguessable URLs rather than being listed anywhere public, so a report is private until you share its link. That also means anyone holding the link can open it: share those links only with people you intend to see the report.

We keep your personal data for as long as your account is open. Your usage records, reports and any replay excerpts on them stay until they are removed, and there are two things worth being exact about, because they differ:

• What you can do yourself, right now. Revoke a connected API key at your provider — that immediately ends anything TokenZip can do with it, whether or not we still hold the encrypted copy. Turn replay off for your workspace, which stops all future replays. Stop sharing a report link, which is what controls who can reach that report.
• What is a request to us, not a button. Deleting a stored credential, a report, a set of ingested usage records, or your whole account is handled by us on request rather than self-service today. Email hello@tokenzip.ai and say what you want removed. We act within 30 days, and we will confirm when it is done.

When we delete, we delete or de-identify the personal data covered by your request — including any connected provider credentials and the prompt and completion text in the affected records — except where we need to retain something to comply with law, resolve disputes, or enforce our agreements. Aggregated or de-identified data may be retained indefinitely. Self-service deletion is a planned change; this section describes what exists today, and it will be updated when that ships.

TokenZip is a business tool and is not directed to children. We do not knowingly collect personal data from anyone under 16. If you believe a child has provided us personal data, contact us at hello@tokenzip.ai and we will delete it promptly.

Depending on where you live, you may have some or all of the following rights over your personal data:

• Access — ask what personal data we hold about you and receive a copy.
• Correction — ask us to fix inaccurate or incomplete data.
• Deletion — ask us to delete your personal data. This is a request by email rather than a control in the product; see section 8 for how it works and how long it takes.
• Portability — receive your data in a machine-readable format.
• Objection and restriction — object to or restrict certain processing.
• Withdraw consent — where processing is based on consent, withdraw it at any time. Replay opt-in is the clearest example: switching it off withdraws it.

If you are a California resident: you have the rights to know, access, correct, and delete personal information, and to opt out of its sale or sharing. We do not sell personal information or share it for cross-context behavioral advertising, and we do not discriminate against you for exercising your rights.

To exercise any of these rights, email hello@tokenzip.ai. We may ask you to verify your identity before acting on a request.

TokenZip is operated from the United States and offered to customers in the United States. Our service providers process and store data in the United States, and that is where your personal data lives.

The Service is not directed to individuals in the European Economic Area or the United Kingdom, and we do not offer it to customers there. If you access it from outside the United States you do so on your own initiative, your data is transferred to and processed in the United States, and you are responsible for compliance with the laws that apply where you are. We would rather say that plainly than publish rights language for a jurisdiction we do not serve.

We may update this policy as the Service evolves. When we do, we’ll post the new version here and update the effective date at the top. If a change is material — a new subprocessor that receives your content, for example — we’ll give you reasonable advance notice by email or in the product before it takes effect.

Questions, concerns, or requests about this policy or your personal data: hello@tokenzip.ai.