Get API key

Shared responsibility

View as Markdown

What the API does with your request, and what stays with your application.

The PrivateMind API runs a model on the request you send and returns what the model produced. It does not run your tools, fetch content for you or act on the answer. Whatever happens after the response arrives is done by your code, so the controls around it live in your code as well.

Your system prompt

The system prompt and the messages in a request are yours. PrivateMind does not add instructions of its own, so the model sees what you sent, formatted by its chat template. The exception is a model built around a fixed persona, where the persona is placed ahead of your system prompt.

The API therefore gives the model no identity, policy or refusal rules for your product. If your users should meet an assistant with a name, a scope and rules, write those into your system prompt. With no system prompt, a model asked about itself answers from its training. It usually gives the name its vendor gave it, and sometimes describes instructions or tools that do not exist.

What you ask a model to produce, and what you do with it, is covered by the Fair Use Policy.

Tools

You run the tools you declare in tools. When the model decides to call one, the API returns the call and stops: finish_reason: "tool_calls" on chat completions, a tool_use block on the Messages API. Nothing runs on the PrivateMind side, whatever the tool is called. Your code decides whether to run it, runs it, and sends the result back. See Tool use.

The decisions about execution are yours to make:

  • Declare only the tools the task needs. A tool that deletes, sends, pays or runs code needs checks in your code, not only a description telling the model when to use it.
  • Validate the arguments against your schema before acting on them. The API does not enforce your schema on the model's output.
  • Ask a person to confirm actions that cannot be undone.
  • Run tools that execute code in a sandbox, without credentials or network access they don't need.

Content you pass in

The model reads documents, web pages, search results, emails and tool results that you add to the conversation the same way it reads your instructions. If that text contains instructions, the model may follow them. This is prompt injection, and no model is immune to it.

Treat anything you retrieve as untrusted:

  • Keep it out of the system prompt. Put it in a user or tool message and mark clearly where it starts and ends.
  • Don't let retrieved text trigger a tool call unchecked. A call that follows a retrieval step gets the same validation and approval as any other, and more when the tool has side effects.
  • Limit what one conversation can reach. A model that reads untrusted email should not also hold a tool that sends email without a person approving it.

Reasoning output

Reasoning models return their chain of thought next to the answer: reasoning on chat completions, thinking blocks on the Messages API. It is unfiltered model output. It can repeat parts of your system prompt and it can make things up. Don't show it to your users unless you mean to. If you don't need it, leave it out of the response with include_reasoning: false on chat completions, or thinking: {display: "omitted"} on the Messages API. See Streaming → Reasoning output.

Moderation

/v1/moderations classifies text you send it. It is one signal for your application to act on, and it has limits of its own. Read them before you rely on it.

Where next