AI Policy

Cheela runs an agent loop: it sends your messages and your capability manifest to a language model, dispatches the tool calls the model returns to your HTTPS endpoint, feeds the results back, and repeats until the model answers or the step budget runs out.

We do not train models, and we do not use your data to train anyone else’s. Model access runs on our own provider credential, which is why you never hand us a third-party API key.

Output is generated by a statistical model. It is not reviewed by a person before it reaches you, and it is not checked for accuracy by us. Two identical requests may produce different answers.

Language models produce confident text that is sometimes wrong. They invent citations, misstate facts, produce code that does not compile, and call tools with arguments that make no sense. This is a property of the technology, not a defect we can remove.

Do not rely on model output for anything consequential without a human checking it. That applies with particular force to medical, legal, financial and safety-critical uses.

• Review output before acting on it, publishing it, or showing it to your own users as fact.
• Disclose to your users that they are interacting with an AI system where they would otherwise reasonably assume a human.
• Design your capabilities defensively. The model chooses which to call and with what arguments — a capability that deletes data or spends money on the model's say-so will eventually do so wrongly.
• Validate capability inputs at your endpoint. Requests are signed so you can verify they came from us, but we do not verify that the model's arguments make sense.
• Comply with the disclosure rules that apply to you. Several jurisdictions now require AI-generated content to be labelled.

The Acceptable Use Policy governs what you may not do. In addition, do not use the service as the sole decision-maker for:

• Medical diagnosis or treatment.
• Legal advice presented to a third party as authoritative.
• Credit, insurance, employment or housing decisions about an individual.
• Anything where a wrong answer causes physical harm.

A person must be accountable for the decision in each of these. Using the model to draft, summarise or research is fine; letting it decide is not.

Inference is provided by third parties through OpenRouter. The specific model is chosen by us and may change — we select it for capability, cost and availability, and we do not guarantee any particular model will remain in use.

Your messages are transmitted to those providers to produce a response. Their handling of that data is governed by their terms, which is why they are named as processors in the Privacy Policy. We do not control their safety filters, and content they refuse to generate will be refused regardless of what this policy permits.

As between you and us, you own the output your executions produce, and you own the input you send. We claim no ownership of either.

Two limits worth stating plainly. Model output may not be copyrightable in some jurisdictions, and we make no representation that it is. And a model can produce text closely resembling its training data, so we cannot warrant that output is free of third-party rights — checking that is your responsibility before you publish or sell it.

Message content is not stored. An execution’s trace records the turn structure of the conversation and nothing anybody wrote. Capability inputs and outputs are stored so you can debug a capability, and those often contain user text rephrased. The Privacy Policy sets this out in full.

Write to hello@cheelalabs.com.