Skip to content
Download
All legal documents

AI and Model Providers

How Cusp uses AI models, what model providers see, our no-training commitment, retention, and the limits of AI output.

Last updated

Cusp is an AI agent. It uses large language models, made by other companies, to read, write and run code for you. This page explains what that means for your data and for what the agent produces. It supplements our Privacy Policy and Terms of Service.

You're working with AI

  • The agent's messages, code, reviews, plans and other output are generated by AI models. They can be wrong, insecure or out of date, even when they sound confident. Review output before you rely on it, especially before you merge, deploy or publish it.
  • The agent can take actions on your computer: edit files, run commands and call tools. What it may do without asking depends on the permission mode, rules and sandbox you set. Crew members can work on a schedule or in response to events, within the budgets and approvals you set for them.
  • Some safety features also use AI. For example, auto mode uses a model to decide whether an action needs your approval, and Hub uses automated analysis to label extensions. These reduce risk but don't remove it.

Which models, and who provides them

You choose the model for each session. Cusp connects to models in two ways:

With your own keys (BYOK). Requests go directly from your device (or your cloud runner) to the provider you chose, such as Anthropic, OpenAI, Google, OpenRouter, Mistral, xAI, DeepSeek, Z.ai, Moonshot, Alibaba, Groq, Together, Fireworks or a gateway you configure, or a local model through Ollama or LM Studio. Your agreement with that provider governs how it handles your data. Cusp doesn't see these requests.

Through the Cusp managed gateway (our keys). Requests go through Cusp's cloud to OpenRouter, which routes them to a provider that runs the model you selected. OpenRouter and that provider act as our sub-processors.

On your device. The iOS app's built-in assistant uses Apple's on-device models, and voice input is transcribed on the device. With a local model, everything stays on your machine.

What model providers see

To do its work, the agent sends the model:

  • your messages and any attachments;
  • the instructions it has loaded, such as project instruction files, skills and the system prompt; and
  • the content it reads while working: files, diffs, command and tool output, web pages, and responses from MCP servers and extensions.

This can include source code, secrets that appear in files or output, and personal data. The agent strips your provider keys from commands' environments and doesn't show expanded secrets to the model, but it can't tell which file contents are sensitive. Use the permission rules and sandbox to keep the agent away from files it shouldn't read, and don't keep secrets in your workspace in plain text.

We don't train on your code

  • Cusp doesn't use your code, prompts, sessions or output to train or fine-tune AI models. This applies on every plan, whether you use your own keys or the managed gateway.
  • On the managed gateway, every request we send through OpenRouter tells it to use only providers that don't collect or train on the data (OpenRouter's data_collection: "deny" setting). Where we send a request directly to a model's developer, such as Anthropic, OpenAI or Google, their commercial API terms apply, and under those terms API inputs and outputs aren't used for training by default. On Team and Enterprise plans, an admin can also require zero data retention: requests then go only to providers that don't store prompts or responses at all, and models without such a provider aren't available. Turning this on or off is recorded in the workspace's audit log.
  • Web search. When the agent uses Cusp's managed search, its query goes through our gateway to our search provider, Perplexity, which under its API terms doesn't train on it. See the Privacy Policy.
  • Dictation. Your voice is transcribed by our speech provider, ElevenLabs, which doesn't use it to train its models. We don't store the audio or the text. See the Privacy Policy.
  • With your own keys, the provider's own terms apply. Most providers' paid API terms say they don't train on API data, but some free tiers do (for example, Google's free Gemini API tier). Check your provider's terms.

Retention

  • Cusp's managed gateway doesn't store the content of requests or responses. We keep usage metadata (model, tokens, cost, timing) for billing and abuse prevention, as described in the Privacy Policy.
  • OpenRouter handles request content and metadata as our sub-processor, under its own privacy policy.
  • Model providers may keep inputs and outputs for a limited time under their own policies, for example to detect abuse and comply with the law; how long depends on the provider. Zero-data-retention providers don't keep them.
  • Synced sessions in the Cusp cloud keep the full transcript, including what the model saw and said, until you delete it.

Usage policies that also apply

When you use a model, you also need to follow that provider's usage policy. Through the managed gateway, the main ones are:

Our Acceptable Use Policy summarises the rules that matter most.

Who owns the output

As between you and us, you own the output, as set out in section 10 of the Terms of Service. Model providers' terms generally assign or don't claim rights in outputs as well. AI output can resemble existing code, including open-source code, and in some countries purely AI-generated material may not be protected by copyright. Check output against the licences that apply before you distribute it.

Questions and concerns

If you think the agent produced something harmful, or a model was used in a way that breaks these commitments, email support@cusp.run, or security@cusp.run for security issues.