What Fire logs. What it doesn't.

1. What is always logged

Every request to a Fire endpoint writes one row to Fire's call log, whether the call succeeds or fails. That row always contains:

• the model and provider that served the call, and the call type (chat, image, speech, transcription, decision, upscale);
• input and output token counts (or the equivalent unit: characters, seconds, images);
• duration, any error, and the price charged to your account;
• the call_id — an opaque handle you can quote when reporting or rating a call.

Why: billing (the price you see on every response is computed from this row), the usage and spend views in your portal, abuse and incident investigation, and the "everything is accounted for" guarantee Fire makes publicly in GET /v1/capabilities. There is no parameter to suppress metadata logging.

2. What is opt-in: prompt and response content

The content of a call — system prompt, messages, image prompt, text to be spoken, transcript, decision state and answers — is stored only when your account has chosen to store it. New accounts start with content logging off.

Retention is enforced, not aspirational. A scheduled job runs daily and blanks the content of any stored call older than your window. The row itself — tokens, price, call_id — stays, so your bill and your call reports keep working.

Per-call override. Even on an opted-in account you can withhold a single call's content by sending the header X-Fire-Content-Logging: off, or "log_content": false in the JSON body. The override can only reduce what is stored; it can never turn storage on for an account that has it off.

Every response tells you what happened. meta.content_logged is true or false on every call, so your integration can verify the policy rather than trust it.

3. What Fire never does with your content

• Sell it, license it, or share it with any third party.
• Train public or third-party models on it.
• Use private content for anything other than showing it back to you (and support or incident work you've asked for).
• Use improve content outside Fire's own evaluation and routing work.
• Infer or store anything from a call whose content you withheld — withheld means the text never reaches disk.

4. Upstream providers

Fire is a pipeline, not a model. Your request is forwarded to the provider behind the model you chose (named on every model page and in every response's provider field). That provider receives the content of the call in order to answer it, and its own data policies apply to what it does next.

As published by the providers themselves — verify with them, since their terms are theirs to change: Eden AI and Venice state that they store no request or response content by default and log only metadata; models reached through Eden AI from OpenAI, Anthropic and Google fall under those labs' API terms, which generally describe retaining inputs and outputs for up to about 30 days for abuse monitoring, with zero-retention arrangements available to the labs' own enterprise customers. Fire does not control those windows and does not claim to.

5. Credentials and the vault

Provider API keys live in Fire's vault and are injected server-side. Your Fire token is a badge, not a key: an agent calling Fire — or a third-party service through the vault broker — never sees, holds or transmits a provider credential. Brokered calls are logged with the credential redacted.

6. Your controls

• Change the setting: client portal → Settings → Content logging. Changes apply to the next call; they do not rewrite the past.
• Shorten retention: same screen. Lowering the window takes effect at the next daily prune.
• Delete stored content now, or close your account: contact us through the portal. Content deletion blanks every stored prompt and response on your account; metadata rows are retained for the period required for billing and audit.
• Check what you're being told: meta.content_logged on every response, and the privacy block in GET /v1/capabilities, which states the current default and modes without authentication.

7. Account data

To run a billing account Fire holds your name, email, prepaid balance, transaction history and the hashed form of your API tokens. Card details are collected and held by Stripe; Fire stores only a customer reference and the card's brand and last four digits. Sign-in is by Google OAuth or an email/password pair stored as a hash.

8. Changes to this policy

This policy is effective 2026-10-13, when content logging became opt-in. Before that date Fire stored call content for every account; content from that period falls under your current retention window going forward. Material changes will be dated here and announced in the portal. The machine-readable summary in GET /v1/capabilities is updated in the same deploy.