Acceptable use.

The short version.

• Only your own accountsPoint the agent at systems and accounts you are authorized to use. Not someone else's inbox, not a site you were asked to stay off, not an account you share credentials into.
• The sandbox is for your workCode runs in an isolated cloud sandbox. Use it for your own compute. Not for mining, scanning, flooding, or building malware.
• Tool approvalsYour Tool Approvals setting decides, and it governs our own tools as well as connectors. The website starts with Skip approvals for eligible actions, including reads inside AGI, web search, page fetch and code in the AGI sandbox. Destructive or unknown actions still ask, as do tools you marked Needs approval; blocked tools stay blocked. You can switch to Ask before every action or read-only auto-approval. Workspace restrictions may require more approval. See the full authority map on agent permissions.
• A block is absoluteA tool you block is refused on the server before it runs. Nothing in the app or the API can override that.
• You can revoke everythingDisconnect a connector, reset a tool's saved permission, or remove a site from the browser allowlist, all listed on the agent permissions page.
• Some uses are off limits everywhereNo sexual content involving minors and no sexually explicit material. No deceiving voters. No content that encourages self-harm. No attacks on critical infrastructure. No passing AI output off as human-made to deceive someone. No harvesting output to copy the models or to build a competitor. Section 02 states each rule in full.
• Breaking these rules costs accessThe ladder is rate limiting, then server-side tool refusal, then account suspension or ban, then termination. Suspension is enforced on every request.

The precise default authority for every tool (what asks, which settings let a tool run without asking, and how to revoke) is on /agent-permissions.

01 · What this covers.

These rules apply to every AGI surface (web, desktop, mobile, the Chrome extension, the VS Code extension, and the CLI) and to all three trust boundaries: Local on-device execution, BYOK using your own provider keys, and Managed Cloud that AGI hosts and meters.

Managed Cloud is in public alpha and open by default. Read this page as the rules for an alpha service, not as a general-availability contract. Where you use BYOK, the provider whose key you supply also applies their own terms to your usage, and those govern that traffic. See “Third-party services and connectors” in the terms.

What the text filter checks, and what it does not. A rule-based filter reads chat messages, image prompts, and video prompts sent through Managed Cloud before any model sees them. The same filter runs when you publish an artifact. It refuses clear requests in five areas: sexual content involving minors; making chemical, biological, radiological, or nuclear weapons; building explosive devices or untraceable or converted firearms; stated plans of violence against a specific person or place; and sexualised or impersonating depictions of real people. A refused request gets this message: “This request was refused because it violates the AGI Workforce usage policy. No model request was sent.” The filter matches patterns of words. It is not a review of everything you write, and it will miss things. It does not look for the other rules on this page, including the election, self-harm, competition, and harvesting rules. Those rest on your agreement to this policy, on the model providers’ own safety rules, and on what we learn from abuse reports. Settings also has a Reduce sensitive content switch, off unless you turn it on, that refuses a short list of explicit or harmful how-to prompts in chat. Scheduled runs, code sessions, and live voice do not pass through the filter. Local and BYOK requests go from your device to the model or provider and never reach these checks.

• Per-user request ceilingChat completions are rate limited per authenticated user: 30 requests per minute at the time of writing. The limiter is fail-closed: if the rate-limit store is unavailable, requests are refused rather than allowed through.
• Per-IP request ceilingA separate, broader pre-authentication ceiling applies per IP address (1,500 requests per minute at the time of writing) so a shared network cannot consume one user's quota. It is also fail-closed.
• Conversation operationsConversation create/update/delete operations are limited to 60 per minute per user.
• API key scopesThe public API supports exactly three scopes: models:read, inference:write, and usage:read. There is no scope that grants connector access, sandbox control, or administrative action. Do not represent otherwise to your own users.
• Key handlingAPI keys are issued to one account. Do not share, resell, sublicense, or embed a key in a distributed client where a third party can extract it. Activity through your key is your responsibility.
• Connector and sandbox ceilingsThe number of connector tools available to your account and the number of concurrent sandboxes you may hold are capped by plan. Do not run multiple accounts to defeat a cap.
• Crawling this siteOur crawler policy is published at /robots.txt and is part of these rules. Application routes are disallowed for all crawlers, and Common Crawl (CCBot) is disallowed entirely.

Do not attempt to evade a limit by rotating accounts, addresses, or keys. Retry with backoff when you receive a rate-limit response rather than retrying immediately in a loop.

• 1 · Rate limitingRequests over a published ceiling are refused. On the security-sensitive limiters, an unavailable rate-limit store also results in refusal rather than access.
• 2 · Server-side tool refusalA tool you have blocked is refused on the server before execution, on both the streaming tool loop and the approval-resume path. This applies to a request from our own app and to a request you construct yourself against the API.
• 3 · Account suspension or banAn account status of suspended or banned is checked on every authenticated request and returns a 403. The check fails closed by default: if the status cannot be read after a retry, the request is refused with a 503 rather than allowed through.
• 4 · TerminationMaterial breach may end your access entirely under the termination section of the terms of service, which also states what happens to your data and which clauses survive.

Appeals. If your account is suspended or banned and you believe it was wrong, sign in and appeal from the notice you see, or use the appeal form with the email address on the account if you cannot sign in. You can also email contact@agiworkforce.com with the subject “Suspension appeal”. Each appeal gets a reference and a person replies. A suspension can be reversed by reinstatement; we will tell you what triggered it unless doing so would compromise an investigation or another person’s safety.

05 · Reporting abuse.

If an AGI agent is acting against you or your systems (unwanted automated traffic, messages from a connected account, or activity you did not authorize), email contact@agiworkforce.com with the subject “Abuse report”. Include the affected domain or account, timestamps with timezone, and any request identifiers, log lines, or message headers you have. You do not need an AGI account to report abuse.

If you found a security vulnerability, use the subject line “Security” and follow the guidance on /security.

AGI Automation LLC, c/o registered agent, 5900 Balcones Drive STE 100, Austin, TX 78731, USA. We may update this policy with notice posted on this page; material changes are recorded on /changelog.