Keys no longer have Standard or Prompt modes. A server prompt is configured independently for Claude, Kimi, Composer and Grok. The model field selects the model; the key and USD balance remain shared.
If any condition is missing, the request proceeds without a server prompt. For Claude, the discount depends on actual application in that request, not the setting label.
| Request condition | Tokens and cache | Discount |
|---|---|---|
| Claude, no server prompt applied | Base cost × 0.75 | 25% |
| Claude, server prompt applied, without PRIVATE | Base cost | None |
| Kimi, Composer or Grok, without PRIVATE | Base cost | None |
| Confirmed current PRIVATE membership, any family | Base cost × 0.75 | 25%, regardless of server prompt application |
- Check the selected surface: Messages, Chat Completions or Responses. A shared address does not mean identical parameters.
- Claude has specific thinking, media and citation behavior when a server prompt is applied; the detailed matrix is on the Messages page.
- Fable 5 and 5.1 work without an applied server prompt. If a prompt would actually apply, a Fable request is rejected before billing. If no suitable prompt is available, Fable works without it.
An ordinary model refusal is returned in the response. Guard does not stop replies with a local refusal or repeated-text detector. Request limits and error handling still apply.
A structured Claude cybersecurity refusal returns the cyber_refusal code:
Claude returns tool_use, the client executes the function, and the client sends the result as tool_result. Guard carries that exchange but does not execute the tool on its own server.
tool_result content | Handling |
|---|---|
String or text | Passed to the model as text |
image | Base64 is supported; a URL is not fetched for a regular key, so download a remote image and send it as base64 |
document | Text sources and base64 PDFs are passed to the model as documents; shared server-prompt media and size limits apply |
search_result | Passed as a client search result and cannot be mixed with other result types in one result |
tool_reference | Passed as a reference to a client-side tool |
is_error: true | Passed to the model as a tool-execution error |
{
"role": "user",
"content": [{
"type": "tool_result",
"tool_use_id": "toolu_…",
"content": [
{"type": "text", "text": "Check completed"},
{
"type": "image",
"source": {
"type": "base64",
"media_type": "image/png",
"data": "…"
}
}
]
}]
}Choose the existing guide for your client. Installation steps are not duplicated here.
anthropic and openaiNode / TypeScriptExamples for both SDKsn8nConnect automationsMigrate from OpenAIMove an OpenAI SDK client to GuardToken cost is calculated at the requested model’s published rate, including cache writes and reads. When eligible, that cost is multiplied by 0.75. Fixed server-tool fees are added afterwards.
- PRIVATE gives 25% off all four families on all present and future keys of a user while their current membership is confirmed.
- Confirmed departure from PRIVATE ends the discount. A lifetime right to return is not current membership.
- If a membership check is temporarily unavailable, the last confirmed state is retained. PRIVATE eligibility does not apply before the first confirmation.
- The new rules do not recalculate historical financial records.
| Setting | What happens |
|---|---|
| Key, address and balance | Stay the same |
| Family | Claude, Kimi, Composer and Grok have independent options and entitlements |
| Request already started | Retains its initial prompt application and price |
| Next request | Uses the current entitlement and option |
| Entitlement expiry | The chosen option is retained; application stops and resumes after renewal |
- Open @subscribe_ai_bot → My API keys and choose the key.
- Open key management and the intended family: Claude, Kimi, Composer or Grok.
- Check the server-prompt entitlement, its expiry and the selected option.
- After changing the setting, send a new request and check actual application in its history.
- The key card shows the option and entitlement for the intended family.
- The account shows actual application and the discount for each request.
- In Usage API, use
familyandendpoint; request records includeserver_prompt_appliedanddiscount_percent. The removedmodefilter returns410 product_modes_removed.
- Server-prompt settings affect subsequent requests. An already started stream retains its original conditions.
- Start a new agent session after changing its server-prompt binding.
- Choose the exact
modelin your client; a server-prompt entitlement does not choose the family for you.
- Reopen the key card if you are viewing an old bot message.
- Check the key, family, entitlement expiry and enabled option.
- Check a new request: a setting change does not affect a stream already in progress.
- If the entitlement is active and the option is enabled but no prompt was applied, ask support whether a suitable prompt is available. Do not send your full API key.