Source: https://respawn.sh/trust/ai

Trust center (https://respawn.sh/trust) 

# AI Usage Policy

EffectiveNov 7, 2026
Last updatedNov 7, 2026
Version3.0.0
EntityRenvo Productions Inc. d/b/a respawn.sh

This AI Usage Policy explains how the AI features of respawn.sh work, what you may and may not use them for, who is responsible for what they do, and the safeguards we apply. It is published by Renvo Productions Inc. d/b/a respawn.sh ("Company," "we," "us," "our") and forms part of the Terms of Service at https://respawn.sh/trust/tos. Capitalized terms not defined here have the meanings in the Terms of Service. The Privacy Policy at https://respawn.sh/trust/privacy explains how we handle the personal data involved.

## 1. Which Features Use AI
These features use AI models (together, the "AI Features"):

- The AI Assistant (Veyra) in the respawn.sh dashboard and, where a Provider has turned it on, in the Embedded Service. It answers questions about a server, reads server information, proposes changes, and carries out the actions you ask for or approve.
- Autopilot, in its Suggest and Protect modes, where the AI Assistant investigates problems it detects on a server you chose (Section 5). Notify mode does not use an AI model.
- Memory, where the AI Assistant saves short notes and recalls them later (Section 7).
- Check on a copy, where a proposed change, such as a plugin update, is tried on a disposable copy of your server before it reaches the real one (Section 6).
- AI summaries and explanations, such as crash explanations, update summaries, anomaly notes on Scrye incidents, the console AI bar, and the store advice and insights in the store tools beta.
Anything an AI Feature produces, including answers, summaries, proposed commands, file changes, and builds, is "AI Output." We label the AI Assistant and AI Output in the Service so that you know when you are dealing with an AI system and not a person.

## 2. The AI Models We Use
We send AI requests through Cloudflare's AI Gateway, which we use to route requests, enforce spending limits, and keep a record of each call's model, size, cost, status, and duration. We configure the gateway so that it does not cache requests and does not store their content.

- Default model. Most AI Features use Anthropic's Claude Haiku 5.5.
- Escalation. When the AI Assistant needs more capability for a task, it may move that task to Anthropic's Claude Sonnet 5.5, which counts 20 times against your AI usage limits (Section 9). Screenshot reviews in Check on a copy use Claude Sonnet 5.5.
- Fallbacks. If our own Anthropic account cannot take a request, Cloudflare may send it to the same Anthropic model using Cloudflare's billing instead. If Claude is unavailable, the request falls back to Google's Gemini models on Google Cloud (Vertex AI). Some functions always use Google models, including the embeddings we use for search and for personal notes, and AI runs for Providers that have not moved to our current host billing.
- Memory processing. Our memory service for server and Organization notes runs on our own Cloudflare account with its own embedding model. It uses an AI model to organize notes, and falls back to an open model (Llama) run by Cloudflare Workers AI. In the dashboard, personal notes are kept in our own database; in the Embedded Service, all notes, including a Panel User's personal notes, are kept in our memory service.
Anthropic, Google, and Cloudflare process this data as our Sub-processors (https://respawn.sh/trust/subprocessors). Under their commercial terms they do not use the data we send to train their models. They may keep it for a limited time, for example to detect abuse, as their terms describe. We may change models or providers; we will update this Policy and the Sub-processors page when we do.

## 3. Acceptable Use of AI Features
You may use the AI Features only to run, maintain, troubleshoot, and improve game servers that you own or are authorized to manage, and only as the Terms of Service and the Acceptable Use Policy allow. You must not, and must not help anyone else to:

- use the AI Features on a server, panel, account, or system that you are not authorized to manage, or to gain access you do not have;
- try to get around an approval, a permission, a spending or usage limit, a redaction measure, a file guard, or any other safeguard, including by instructions hidden in server files, logs, chat, player names, plugin descriptions, or other content the AI Assistant reads ("prompt injection");
- use AI Output or AI actions to attack, overload, or disrupt any server, network, or person, including through denial-of-service, crash tools, or cheats aimed at others;
- ask the AI Features to produce malware, credential-stealing tools, exploits for systems you do not own, or instructions to break the law;
- use the AI Features to harass, threaten, doxx, or discriminate against anyone, including players, or to take abusive moderation actions such as mass kicks or bans without a legitimate reason;
- deliberately give the AI Features special categories of personal data (such as health, biometric, or government identification data), payment card data, or personal data of children, beyond what is unavoidably present in normal server files;
- use the AI Features to process protected health information or to provide medical, legal, or financial services;
- extract, copy, or harvest AI Output, our knowledge base, or our prompts in bulk, or use them to build or train a competing AI system;
- misrepresent AI Output as written by a person where that would mislead others, or remove labels showing that content was produced by AI where the law requires them; or
- resell, sublicense, or provide access to the AI Features to anyone outside your Organization, except as a Provider through the Provider Platform.
We may refuse a request, stop a conversation, or pause an AI Feature when we detect a likely breach of this Section.

## 4. Authorization of Actions
The AI Assistant may read information about a server without asking first. What it may change, and how, depends on the action:

- Always needs approval. Console commands; starting, stopping, restarting, or killing a server and other power actions; restoring a backup; reverting a file; and kicking or banning a player always require a person with the right permissions to approve that specific action, as do the other actions the Service shows as needing approval, such as scheduling tasks, testing an update on a copy, and changing proxy settings. No setting, Autopilot mode, or Provider setting can approve them in advance, except update protect mode where you select it (Section 5). Each approval covers one action and expires if it is not used within a short time.
- Approval after reading server content. Any change the AI Assistant proposes after it has read server content in the same conversation needs approval, so that instructions planted in server content cannot trigger changes by themselves. The AI Assistant must read a file before it can change it.
- Other changes. Other changes, such as editing a file, creating a snapshot, changing alert rules, resolving an incident, or starting a profiling session, may run without a separate approval when a person with the right permissions asks for them. In the Embedded Service, a Panel User can choose "Ask every time" to approve every change, within the limits the Provider sets.
- Guarded changes. Where we can, a change is made in a way that can be put back: we keep the previous version of a file, take a snapshot before risky changes, and put a change back if writing it fails.
Who may authorize. An approval or request counts only if it comes from a person whose role in the Organization (or, in the Embedded Service, whose Panel permissions) allows that action. Approvals can be given in the dashboard, in the Embedded Service, or, where the Organization has connected our Discord bot, with an approval button sent through Discord. In a Discord server, the approver must be a linked respawn.sh member who holds both the mapped Discord role and the respawn.sh permission the action needs. In a direct message, only the linked Discord account it was sent to can approve, and its respawn.sh member must hold the permission. You are responsible for who you give those roles and permissions to, and for the Discord roles you map to them.
Your responsibility. You choose what Autopilot may do. Every action you ask for or approve, and every action Autopilot takes within the settings your Organization chose, is treated as an action taken on your instructions, as if you had taken it yourself through your panel or the dashboard. Read the proposed command, change, or plan before you approve it.
Our responsibility. We are responsible if Autopilot acts outside the settings your Organization chose, if our software malfunctions, if an approval rule fails and an action runs without the approval it needs, or if our system carries out something other than what was asked for or approved. A resulting loss of data is a Data Incident under Section 12.4 of the Terms of Service (see also Sections 12B.5 and 12B.6).

## 5. Autopilot
Autopilot is off unless an owner or administrator of the Organization, or a member they give the Autopilot permission, turns it on for a server. In Notify mode, the Service watches for problems and sends notifications without using an AI model. In Suggest mode, the AI Assistant also investigates with read-only tools and suggests actions for a person to approve. In Protect mode, it may also take, without asking, only the actions you selected from a short list (and, where you turn it on, the plugin updates described under update protect mode below): creating a snapshot, opening or adding notes to an incident, writing to the server timeline, starting or stopping a profiling session, and sending notifications. Except for update protect mode (below), Autopilot never takes an action that always needs approval (Section 4).
Herd actions. When our fleet analysis has found a plugin version harmful on other servers, Autopilot may suggest rolling that plugin back to its previous version from your own snapshot and setting aside the flagged plugin file. A herd action never runs automatically, in any mode: it runs only if a person with the right permissions approves it, and it is recorded on the server timeline and in the audit log. We do not apply fleet-wide changes to customer servers without that approval.
Update protect mode. Where we have turned it on for your Organization and you turn it on for a server, the Service may apply plugin updates (security fixes only, or every safe update, as you choose) without asking, but only after the update has passed a Check on a copy, through a guarded change that takes a checkpoint first, restarts the server once, and is checked for 15 minutes afterwards, within a daily limit. You are told how each update went. Autopilot has a daily limit on investigations and uses your AI allowance, and it falls back to notifications only when either runs out. Autopilot does not catch every problem (Section 8). We may pause Autopilot for any server, Organization, Provider, or everyone at any time, and we tell the affected Organizations and Providers when we do.

## 6. Check on a Copy and In-Game Screenshots
Copies. Before some changes reach your server, we may boot a disposable copy of your server from its latest snapshot on our infrastructure (a Google Cloud job), next to an unchanged copy, and compare the two. The copy has no network access while it boots, holds no credentials, and is destroyed when the check ends. A passed check is not a guarantee that the change will work on your live server.
In-game screenshots. During a check, we may take screenshots inside the copy and have Claude Sonnet 5.5 review them for visible problems. Screenshots are images and are not redacted. They are deleted after 30 days, and finished check runs after 90 days.

## 7. Memory
Where memory is on, the AI Assistant may save short notes about your servers, your Organization, and how you like to work, and use them later. Notes that the AI Assistant proposes are refused before they are saved if they contain personal identifiers, links, or instructions. Where memory is on, the redacted text of finished conversation turns from members who manage memory (the question and the answer) and of Autopilot findings that report a problem is also stored in the memory service for that server or Organization, so that notes can be drawn from it. This text is kept until the server or Organization is deleted. Notes drawn from it are deleted 12 months after they are created, and notes about incidents 6 months after they are created. You can view and delete your personal notes, members with the Autopilot permission can view, edit, and delete notes about the Organization's servers, and the Organization can turn memory off. Notes added directly are deleted no later than 12 months after they were last used, and in any case 24 months after they were created; notes drawn from stored conversation text are deleted as described above; and all of them are deleted when the server, Organization, or account they belong to is deleted, as the Privacy Policy describes. Deleting a single note removes it from use at once; the Privacy Policy explains when earlier versions are erased. Memory for servers and Organizations runs on a memory service we operate on our own Cloudflare account. Each server's and Organization's notes are kept under their own label, and every request is limited to a single label, so notes are never shown to other customers. Fleet-wide notes that we draw from aggregated, de-identified data across many servers are kept under a separate shared label and never contain any one customer's notes.

## 8. Output Limitations
AI OUTPUT IS GENERATED AUTOMATICALLY AND CAN BE WRONG, INCOMPLETE, OUTDATED, OR UNSUITABLE FOR YOUR SERVER, EVEN WHEN IT SOUNDS CONFIDENT. In particular:

- the AI Assistant can misread logs, configuration, or metrics, and can propose a change that makes things worse;
- answers may rely on public documentation for plugins and server software that is out of date or wrong, which we do not control;
- Autopilot can miss a problem, report a problem that is not there, or take a selected action that turns out to be unnecessary;
- a passed Check on a copy does not prove that a change is safe on the live server, which can differ from its snapshot; and
- builds, summaries, and advice (including store advice and insights) are suggestions, not facts or guarantees of any result.
AI Output is not legal, financial, security, or other professional advice. Use your own judgment. Keeping your own copy of important data is good practice, but it isn't a condition of our responsibility (Terms of Service, Sections 7.6A and 12.4). The limitations of liability in the Terms of Service apply to AI Features and AI Output, and so do our responsibilities in Sections 12.4, 12B.5, and 12B.6 of the Terms of Service (Section 4).

## 9. Usage Limits and Charges
AI use counts against your plan's AI usage limits, which are measured in Claude Haiku 5.5 token equivalents over a rolling 5-hour window and a rolling 7-day window and shown in the dashboard, and against any AI Credits available to you. Escalated tasks use the limits faster: a Claude Sonnet 5.5 token counts as 20. When a window runs out, the AI Features pause until enough earlier use falls outside the window, or until you buy a reset where one is offered. A reset is not refundable once it has been applied, except where the law requires otherwise. It is withdrawn if its payment is refunded, and it cannot be used while a dispute about its payment is open. The Subscription and Billing Terms at https://respawn.sh/trust/billing describe the limits for each plan, resets, and AI Credits, and the Refund and Cancellation Policy at https://respawn.sh/trust/refunds describes refunds. In the Embedded Service, AI use is billed to the Provider, never to Panel Users.

## 10. Safeguards
We apply these safeguards. They are automated and can fail, so do not rely on them as your only protection. If an approval rule fails and an action runs without the approval it needs, we are responsible, as Section 4 says.

- Redaction. Before server content is sent to a model by the AI Assistant or the console AI bar, or stored in a transcript, our software replaces credentials and secrets, and replaces IP addresses, email addresses, player IDs, and player names with placeholders. Screenshots reviewed in Check on a copy are images and are not redacted (Section 6), and neither are screenshots you attach to a chat, which we keep for 72 hours.
- File guard. The AI Assistant refuses files that hold secrets and does not open files that usually hold player identifiers unless you ask for that file by name.
- Approval rules. The approval rules in Section 4, enforced in our database as well as in our software.
- Spending limits. Per-call reservations, daily spending caps, and the limits that the payer or Provider sets.
- Audit trail. We record AI actions, approvals, and who approved them, so an Organization can see what was done.
- Kill switches. We can turn off the AI Assistant, Autopilot, memory, or any capability for anyone at any time, and we tell those affected when we do. Removing an AI feature included in a paid plan follows Section 12B.9 of the Terms of Service.
- Testing. Before we change a model or a version of the AI Assistant, we may replay redacted records of past conversations against the new version to check that it behaves safely. This testing does not train any model.

## 11. Training
We do not use your conversations, memories, server content, screenshots, or other customer data to train or fine-tune any AI model. We plan to fine-tune a model for the AI Assistant in the future. Before that starts, we will update the Privacy Policy and this Policy, and the training will be off unless you opt in, with the other safeguards in Section 12B.8 of the Terms of Service.

## 12. Reporting Problems
If AI Output is harmful, an action went wrong, or you think someone is misusing the AI Features, contact support@respawn.sh. Report security issues as described on the Security page at https://respawn.sh/trust/security.

## 13. Changes
We may update this Policy as provided in Section 23.1 of the Terms of Service.