Privacy Policy

We collect the email address you use to create an account. When you use the AI Hub features, we store metadata about your requests (token counts, model used, timestamps) to provide usage analytics. The chat feature (persistent dashboard conversations) has been discontinued. Any conversation history stored while that feature was available is retained solely so we can fulfill your data-export and deletion requests. This historical data belongs to you — we do not use your conversation content to train models, improve our algorithms, or for any purpose other than providing the service to you and running security scans you have enabled. API keys you provide for third-party services (OpenAI, Anthropic, Gemini) are stored in Google Cloud Secret Manager and are never written to our database or logged. Our database retains only a masked display hint. See our Security page for full details. When you register an agent, we store the agent name and configuration. Push notification device tokens are stored to deliver approval notifications.

We use your email address to authenticate you and, if you opt in, to send you notification emails about approval requests requiring your attention. Usage metadata is used to provide the analytics dashboard. We do not sell your data, use it for advertising, or use it to train AI models. When you enable runtime security features (CleverGuard, content scanning, agent drift detection), we process your AI inputs and outputs in real time to detect threats such as prompt injection, PII exposure, secret leakage, and unauthorized tool usage. This processing is performed solely to enforce your security policies. Scan verdicts, categories, and scores are logged to your audit trail. When a scan flags PII or a secret, the finding also keeps a short excerpt of the matched text so you can triage it; if you have enabled zero-knowledge I/O encryption that excerpt is encrypted with your key before it is stored, and we cannot read it. The security scanners and governance models that run on Clevername's own compute (the CleverGuard ML classifier and deep-scan judge, ScopeGuard mandate extraction and scope embeddings, ClaimGuard claim extraction, questionnaire extraction, the agent-questionnaire prefill button, and false-positive report review) run on Clevername-operated, self-hosted models. The questionnaire prefill button was the last lane to move onto our own hardware, on 14 September 2026. If that hardware is unreachable the feature is skipped or returns an error, never handed to a third-party provider. The carve-outs, all described in our DPA: - Scout, our support assistant, sends the questions you type into the support widget to Anthropic, and OpenAI embeds support and memory-search queries. Neither touches your agents' prompts, responses, or logs. - The agent-questionnaire prefill button in the new-agent wizard sends the description you write and any workflow file you paste to our own self-hosted model, and returns suggested answers. It runs only when you press it, and no third-party provider sees that text. If our hardware is unreachable the button returns an error and you fill the questionnaire yourself.

Historical conversation data from the discontinued chat feature is retained only to satisfy export and deletion requests; you can request deletion of all your data at any time. Audit logs and security events are retained according to your plan: 7 days on Free, 30 days on Pro, and 365 days on Team. Enterprise retention is set by contract with no fixed platform expiry. Stored I/O and tool-call content (encrypted or not) is purged after 90 days by default. PII and secret-detection (DLP) findings, including their excerpts, are purged after 365 days by default; organizations can shorten this. Tool arguments parked for a human approval are deleted when the action runs or the request expires. Usage analytics are retained for 12 months. Account data is retained until you delete your account. If you have enabled zero-knowledge I/O encryption, the records our servers can still read are deleted on a shorter clock than the 90-day default: a resolved approval record (its parked arguments and, once run, its result) and a finished asynchronous tool execution are removed once they are 7 days old, by a sweep that runs daily. Files you store through the agent file tools are not on a timer — they are your documents and stay until you delete them or delete your account. You may request an export of your data or erase all of your data by contacting us at privacy@clevername.net. Account deletion is available in dashboard settings. Erasure removes your conversation content, projects, API keys, and analytics. Audit logs are anonymized rather than deleted to preserve the integrity of organizational security records.

When you route AI requests through Clevername (via the dashboard chat or the gateway), your requests are forwarded to your chosen AI provider using your own API keys. We process request and response content in transit to apply your security policies (content scanning, tool drift detection, guardrail enforcement). Content scanning happens in memory, on plaintext, before anything is stored. What is stored afterwards depends on two settings: - Platform-wide default (metadata only): the agent I/O log and the tool-call log record timestamp, model, token count, latency, verdicts, and a content hash — not the prompt, response, or tool text. Storing a copy of prompt and response text is off for everyone unless Clevername enables it platform-wide; there is no per-organization switch. - Per user (zero-knowledge encryption): you can turn on I/O encryption under Settings → Privacy. Your browser derives a key from your passphrase and generates an RSA keypair; only the public key reaches us. From then on every stored copy of your agents' prompts, responses, tool arguments and results, threat-finding excerpts, system-prompt previews, false-positive report text, and ClaimGuard evidence is encrypted with a per-record AES-256-GCM key wrapped to your public key. We keep the ciphertext and cannot decrypt it; there is no escrow copy and no support override. Scanning still happens on the plaintext in memory before encryption, so this is encryption at rest, not end-to-end. One gap we have not closed, stated plainly: on a ClaimGuard evidence row only the evidence span is encrypted to your key. The claim text, its subject and the verdict reason sit next to it in the clear, and they usually restate the same sentence the span came from. Those fields are readable by us. They are not zero-knowledge and we do not describe them as such. Four exceptions are deliberately readable by our servers, because the feature cannot work otherwise: (1) the arguments of a tool call that is parked waiting for a human approval are encrypted with Clevername's platform KMS key so the approved action can run, and are deleted the moment it runs or the request expires — the record of the approval itself is deleted 7 days later; (2) the result of an asynchronous tool execution is held under the same platform KMS key, tied to your account and that execution, so the agent that started the call can poll for it — with encryption on it is deleted 7 days after the execution finishes, rather than at the 90-day default; (3) the contents of files your agent reads are held under the same platform KMS key, tied to your account and that file, so the agent can receive the body it asked for — the agent runs on our servers and has no access to the private key in your browser, so it could not open a zero-knowledge blob — unlike the other two, these are not deleted on a schedule, because they are files you chose to store; they stay until you delete them; (4) content you explicitly donate for training (see "Training data"). For all four, someone with our KMS key and database access can read the content — the envelope means a database dump on its own is not enough. We do not use any of these four for anything other than delivering the feature that requires them. They are not read by staff as a matter of routine, are not used to train or evaluate models, are not used for analytics or product research, and are not shared with third parties. The only routine access is the one the feature performs on your behalf. The content hash stored with each I/O record is keyed with a server-held key scoped to your organization, so a stolen copy of the database is not a lookup table for short prompts. It is not zero-knowledge: Clevername staff who can also read that key could still test guesses against it. Diagnostic-only forensic captures are not retained in a form Clevername can read when your encryption is on, and our short-lived Redis caches hold no content for users with encryption on. Clevername is a BYOK platform for customer inference. AI requests you make are routed using your own provider API keys — we do not supply managed LLM compute for your inference. Clevername does run its own self-hosted compute for governance checks (see "How we use your information"); those models process content only to enforce your security policies and are not customer-facing inference. The same privacy protections apply to all requests regardless of which provider you use.

Clevername relies on the following service providers (sub-processors) to operate the platform. All of them process data in the United States: - Cloudflare — DNS, CDN and web application firewall for clevername.net. Cloudflare terminates the HTTPS connection at its edge and forwards the request to our origin over a second connection, so every request you make to Clevername — including the ones carrying agent prompts and tool arguments — passes through Cloudflare's network in the clear on that hop. Cloudflare does not store request bodies. Cloudflare's edge is global rather than US-only. - Google Cloud Platform — hosts the Clevername API (Cloud Run) and stores your encrypted provider keys in Secret Manager. Also runs our self-hosted CleverGuard Tier-2 classifier, Tier-3 judge and internal governance models on GPU Cloud Run in us-east4 and us-central1. - Vercel — hosts the clevername.net web application and its edge network. - Supabase — database hosting and user authentication; its infrastructure runs on AWS (us-east-1). - Amazon Web Services — hosts the GPU box that runs ClaimGuard's claim extractor (reached over a Cloudflare Tunnel), and, through Supabase, our database. CleverGuard's Tier-2 classifier moved off AWS onto Google Cloud Run on 3 September 2026, and the Tier-3 judge on 5 September 2026. - Stripe — payment processing and subscription billing. Card details are entered directly with Stripe and never touch Clevername's servers. - Upstash — serverless Redis used for rate limiting and short-lived caching. Holds rate-limit counters keyed by IP address or account, and for users who have not enabled I/O encryption, tool-call results cached for up to five minutes. Nothing content-bearing is cached for users with encryption on. - SendGrid — transactional email delivery (account notifications and security alerts). - Anthropic — answers for Scout, our in-app support assistant. Scout sends the questions you type into the support widget, plus the public help articles it retrieves, to Anthropic on Clevername's own key. Scout never has access to your agents' prompts, responses, or logs. - OpenAI — embedding generation: our help documentation, the questions you type into Scout, and memory-search queries for users who have not connected their own OpenAI key. Your agents' prompts, responses, and tool content are not sent to OpenAI on Clevername's keys. - Sentry — application error monitoring. Receives stack traces and request metadata (route, status, timing, user and org identifiers) when an error occurs. Request bodies, including prompts and responses, are never sent. The complete list, with what each provider receives, is in Article 8 of our Data Processing Agreement. Your API keys for OpenAI, Anthropic, and Google are transmitted directly to those providers when processing requests. We do not share your data with any third party except as necessary to route your AI requests to the provider you select.

We never train models on your agents' prompts or responses unless you explicitly donate them. From Settings → Privacy you can pick a time window and donate the I/O records inside it to improve CleverGuard. If your I/O is encrypted, your browser decrypts the donated records with your passphrase before sending them; donated records are then stored readable by Clevername for training and are visible to Clevername engineers. Donations are never shared publicly or with third parties. Nothing is donated unless you complete that flow.

Clevername uses only strictly necessary cookies — cookies that are essential to operate the service. Specifically, we set session cookies issued by Supabase (our authentication provider) to keep you logged in. These cookies are required for the service to function and do not track you across other websites. We do not use advertising cookies, analytics cookies, or any third-party tracking technologies. No consent is required for strictly necessary cookies under applicable law, but you may disable cookies in your browser settings. Doing so will prevent you from logging in.

When you connect a Google account to sign in, Clevername accesses only the Google user data required for that specific purpose. Sign-in: We request your email address and basic profile information to create and identify your account. Google Drive (discontinued): The Google Drive file-attachment integration has been removed. Clevername no longer requests, reads, or stores Google Drive files, and no longer requests Drive scopes at sign-in. Any file references stored while the integration was available are retained only to fulfill export and deletion requests. Clevername's use of Google user data complies with the Google API Services User Data Policy, including the Limited Use requirements. Specifically: • We use Google user data only to provide the features you explicitly enable — we do not use it for advertising, profiling, or any purpose unrelated to the feature you authorized. • We do not share Google user data with third parties except as necessary to deliver the service (for example, to authenticate your sign-in session). • We do not allow humans to read your Google user data unless you explicitly request support and grant access, or it is necessary for security purposes such as investigating abuse. • We do not use Google user data to train machine learning models.

All data is transmitted over HTTPS. Your HTTPS connection terminates at Cloudflare's edge, not at our servers; Cloudflare then opens a separate encrypted connection to our origin. That is standard for a CDN, and it means the transport is not encrypted end-to-end between you and our compute. API keys are encrypted at rest in Google Cloud Secret Manager. Authentication tokens are stored securely using platform-appropriate mechanisms. We conduct regular security reviews and follow industry best practices for data protection. Our infrastructure runs on Google Cloud Platform with strict IAM policies and no public endpoints.

You may access, correct, or delete your personal data by contacting us at privacy@clevername.net or through dashboard account settings. To request a complete export of all data we hold about you, contact privacy@clevername.net. If you have enabled I/O encryption, the export includes your encrypted I/O records as ciphertext; you hold the only key that can decrypt them. Upon account deletion, all your data is permanently removed within 30 days.

We may update this policy from time to time. We will notify you of significant changes by email. Continued use of the service after changes take effect constitutes acceptance of the new policy.

For privacy questions or data requests, contact us at privacy@clevername.net.