Is it safe to send your code to a hosted model API?

Updated 13 November 2026

When a coding agent calls a hosted model, the files it reads and the prompts it writes leave your machine and reach someone else's servers. Whether that's acceptable depends on the provider's rules and on what your code is. Here's what to check, and what NeverMetered does.

What to check about any provider

  • Whether your content is used for training. The terms should say no outright, not leave you to find a setting.
  • Whether it's stored, and for how long. Look for whether prompts and outputs are written to logs or a database, and whether there's a retention period.
  • Who can see it. Ask if staff can read your content, and whether third-party model providers receive it.
  • Where it's processed and by whom. Hosting providers and sub-processors should be listed in the privacy policy.
  • Whether your employer or client allows it. For work code this comes before everything else, and a provider's privacy policy doesn't override your own contract.

What NeverMetered does

What each step sees and keeps, from our privacy policy.

This summarises our Privacy Policy, which has the full detail and is the source of truth if the two ever differ. We never use your prompts or the model's outputs to train models. Requests are processed in memory, and content isn't written to our database or our logs. Logs record metadata such as ids, timings, token counts and the number of messages.

To make follow-up turns faster, the inference servers keep recent conversation state in GPU and system memory. That cache isn't returned to anyone as output, it's overwritten as new requests arrive, and it's lost when a server restarts. If you use the OpenAI Responses API without store: false, the server also keeps the response in memory so you can continue the conversation. Only your account can reach it, it's never written to disk, and it goes when it's evicted, deleted or the server restarts.

Our team can see request metadata for support and abuse prevention, but there's no store of your content to look at. Which companies help us run the service, including the hosting providers that operate the physical machines, is listed in the policy's Who we share data with section. That list is kept current, so check it rather than relying on this page.

What that doesn't cover

A privacy policy says what a company does. It isn't an independent audit, and we're a small provider, not a large vendor with enterprise contracts. If your organisation needs a formal compliance statement or a data processing agreement, ask them before you send work code to any small provider, including us. How it runs is written so you can check what we claim.

Habits that help with any provider

  • Keep secrets out of the repository and out of files the agent reads: .env files, private keys and tokens.
  • Use a separate API key for each project, so you can revoke one without touching the others. See one key per project.
  • Use a hosted API for projects you're comfortable sharing and run a local model for code you aren't. See the local-versus-hosted guide.
  • Re-read the terms when they change. Ours are dated, and paid plans have notice commitments for changes.
Does NeverMetered train on my code?
No. We never use your prompts or outputs to train models.
Are my prompts stored?
No. They're processed in memory and not written to our database or logs. We keep request metadata such as token counts and timings for your usage dashboard.
Can I use it with private repositories?
That's your decision and, for work code, your employer's. Our policy is to not store or train on content, but we don't claim a compliance certification. Check with whoever owns the code.
Sources