> ## Documentation Index
> Fetch the complete documentation index at: https://docs.pureframe.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Agent Security

> What an agent can and can't do with your library, and how to scope it safely.

An agent is only as trustworthy as the credential you give it. Pureframe AI's controls for agent access are the same ones that control any API key — there's no separate agent-specific permission system.

## Use a read-only key

Create an API key with `permissions: "read_only"` specifically for agent integrations that only need to search (see [API Keys](/get-started/api-keys)). A read-only key can call `search_videos`, `list_collections`, `get_collection`, and `get_video`, but not `upload_video` or any write/delete operation — even if a prompt injection attempt tries to talk the model into it.

```bash theme={null}
curl -X POST https://api.pureframe.ai/v1/api_keys \
  -H "Authorization: Bearer <session-token>" \
  -H "Content-Type: application/json" \
  -d '{ "name": "Claude Code — read only", "permissions": "read_only" }'
```

## What's scoped, and what isn't

An API key is scoped to your account — it can access every collection and video you own, and nothing outside your account. There's currently no way to restrict a key to a subset of collections (a "this agent can only see the Marketing collection" allowlist doesn't exist yet). If you need to give an agent narrower access than your whole library, the practical option today is a separate Pureframe AI account holding only the collections that agent should see.

## Revocation is immediate

If an agent integration is compromised or misbehaving, revoke its key from the [API Keys dashboard](https://platform.pureframe.ai/settings/api-keys) — in-flight and future requests using it are rejected within seconds. Keys are stored as a SHA-256 hash server-side, so Pureframe AI can't recover or display the plain secret after creation, only revoke it.

## Rate limits apply the same way

Agent tool calls share the same per-key rate limits as direct REST calls — 30 requests/minute for `search_videos`. A runaway agent loop hits `429`s, not unlimited API usage. See [Rate Limits](/production/rate-limits).

## Remote MCP authentication

The hosted remote MCP endpoint (`https://mcp.pureframe.ai`) authenticates the same way as the REST API — a Bearer token in the `Authorization` header, passed through by your MCP client's config. There's no separate remote-specific credential type.