Acceptable use policy

This policy applies to every call you make through the service, whether from the console, the API, or any third-party client.

You are responsible for all requests made with your account or keys, including those you hand to colleagues, embed in your own product, or issue from automation.

Do not use the service to commit crimes or to generate or distribute unlawful content.

Do not infringe intellectual property, reputation, likeness or other legal rights of others.

Do not generate sexualised content involving minors. There is no exception to this — not research, not testing, not art.

Do not forge legally effective documents, identity papers, financial instruments or official records.

Do not generate actionable instructions for violence, terrorism or self-harm.

Do not generate synthesis or manufacturing guidance for weapons, explosives, dangerous chemicals or biological agents.

Do not use the service to harass, threaten, defame or dox a specific individual.

Do not generate content designed to deceive, including impersonation of real people or institutions and fabricated news or reviews.

Do not submit other people's personal data to a model without their consent. If your product processes end-user data, establishing a lawful basis is your responsibility.

Do not use the service for unauthorised facial recognition, behavioural profiling or identity inference.

Worth stating explicitly: as the party submitting content, you carry obligations of the same nature as ours. That we do not retain your content does not make it safe to submit data that should never have left your systems.

Do not attempt to bypass quota, metering or billing, including falsifying usage reports or racing concurrent requests to spend the same allowance twice.

Do not load-test, scan or penetration-test the service beyond normal use without our prior written permission.

Do not reverse engineer or decompile our client tooling, or attempt to reach other users' data, keys or usage.

Do not operate the service as a proxy egress for third-party traffic unrelated to you.

Do not resell or rent the service, or share accounts or API keys outside your own organisation. Providing AI features to your own product's users is fine — you are calling us from your product. What is not fine is selling our allowance itself onward as a product. The distinction is whether you remain accountable for what your end users do.

Do not circumvent regional restrictions. Some models have not completed mainland-China filing and are not offered to users in mainland China; accessing them by concealing your location or using a proxy is at your own risk (see section 4 of the Terms).

Do not breach the usage policy of the upstream provider behind the model you are using. Links to those policies are on each model's detail page.

An honest policy has to answer this. We commit to never retaining or reading your request content (Privacy Policy, section 3) — so how would we ever find a violation?

Mostly, not by looking at content. Three things are available to us. First, anomalies at the metadata layer: call volume and source-IP distribution that clearly fits proxy egress, or concurrency timing that fits quota circumvention. Second, violation reports from upstream providers — they run their own detection on the content they receive and tell us which call tripped it, which we map back to an account. Third, third-party reports and demands from law enforcement.

This makes our enforcement limited, and deliberately so. A gateway that could detect every violation would necessarily be a gateway that reads everything. We chose the ceiling on the former over the floor on the latter.

We will not make an exception and read your content in order to enforce this policy. In almost every case we technically could not: the history does not exist and cannot be recovered. One exception belongs here, so the sentence above is not read as a stronger promise than it is: when you use `previous_response_id`, that session's messages live in our database for up to 5 hours (the same exception documented in Privacy Policy section 3), and during that window they are technically readable. We do not read them, and a scheduled job deletes them at expiry; outside that window the content genuinely does not exist. Where a specific call must be examined, only you can produce it, knowingly and voluntarily.

We escalate in stages: notice (email, no effect on service) → rate limit (reduced concurrency, service continues) → suspension (paused, restorable on appeal) → termination (account closed).

Except for serious cases under sections 2 and 3, we give notice before acting, naming what to fix and a reasonable deadline. We do not close accounts without warning.

On suspension or termination, unconsumed paid balance and unused quota are handled under the Refund Policy. Only termination for reselling, circumventing regional restrictions or serious unlawful conduct forfeits a refund.

Where conduct may be criminal, we report it to the competent authority as the law requires and preserve the relevant records.

Any enforcement action can be appealed within 30 days to appeal@agiplan.dev, with whatever material you consider relevant.

We respond within 5 working days, and the review is carried out by someone not involved in the original decision. A successful appeal restores service and credits the quota lost while suspended.

Where an action was based on a metadata inference and you offer a plausible explanation, we restore your service first and continue investigating afterwards.

Material changes are announced 30 calendar days ahead. To address newly emerging abuse or a legal requirement we may add prohibitions with immediate effect, but we will say why.

If you do not accept a change, you may cancel before it takes effect and receive a pro-rata refund.