Privacy Policy

Inference Ghost (agents.inference.ai) is a per-user AI agent platform. Each customer runs a dedicated VM and connects messaging channels (Telegram, Discord, WhatsApp, Lark, Google) to their agent. This page describes what we store, where it goes, and why.

When you sign up we store your email, display name, and (if provided) profile image. We also store your password (hashed) and, if you sign in with Google, encrypted Google OAuth tokens. Each session records the IP address and browser you signed in from. Your login session is shared across Inference AI apps via a cookie scoped to inference.ai. Authentication is provided by Better Auth (auth.inference.ai); account data lives in managed Postgres.

When you connect Discord we store the minimum routing data in our database (D1):

• Your own bot: your bot token.
• A managed bot from our pool: the assigned bot's token, the server you added it to, and a reference to the pooled bot. Connecting a managed bot uses Discord OAuth with the bot and identify scopes; we do not store your Discord identity.
• Both modes: the Discord user IDs your bot is allowed to answer, and the list of servers your bot is in.

Discord messages go directly from Discord to your VM and are handled by the agent running there — message content does not pass through our own servers. Generating a reply sends that content to a model provider (see below). We keep short activity summaries for your dashboard, but not full transcripts.

Telegram: your bot token, the user IDs we route to your VM, and temporary pairing tokens, stored in our database (D1). WhatsApp: we store only a session reference (the linked account's ID and pairing QR) in D1 — the actual session credentials live on your VM and in your encrypted VM backups. Disconnecting clears what we store; a Telegram bot token stays valid until you revoke it, and a WhatsApp device link stays active until you unlink it from your phone.

If you connect Lark/Feishu, we store your Lark app ID and app secret in our database to route messages. If you connect Google, we store the Google account email; the browser on your VM stays signed into that account so your agent can act on your behalf. You can disconnect either from the dashboard.

Your VM runs your agent (Hermes, Claude Code, OpenClaw). Files, conversation history, and tool outputs are created and stored on your VM. To generate responses, your agent sends prompts and conversation context to third-party model providers (see "How your agent generates responses"). Durable VM state — conversation stores and, if connected, your WhatsApp session — is also replicated to encrypted, per-customer backups in Cloudflare R2. We do not read, mine, or train on this data.

To answer, your agent sends your prompt together with recent conversation history and tool outputs to a large language model. Model calls are routed through our gateway to the provider behind the model you select — for example Anthropic, OpenAI, Google, DeepSeek, Moonshot, Zhipu, Alibaba, xAI, or Mistral (over 100 models are available). That content leaves your VM on every turn and is processed by the chosen provider under its own terms. Some models are operated by providers outside the United States, including in China; you choose the model, and therefore which provider — and which country — receives your content. The gateway records usage totals (token and request counts, cost) per account for billing, but does not store your prompts or your agent's replies. We don't train on your conversations and we don't authorize the gateway or providers to; each provider's own retention and training practices are governed by its terms.

Channel metadata is kept for as long as the channel remains connected. When you disconnect a channel, we clear its credentials from our database within seconds; some routing identifiers (such as the user IDs we route to your VM) are retained until your VM is destroyed. A managed pooled bot leaves your server on disconnect; a bot you connected from your own account stays in your server until you remove it. WhatsApp and Telegram sessions remain valid on the provider's side until you unlink or revoke them there. To delete your account, email hello@inference.ai — we destroy your VM, revoke its model-gateway key, delete its externalized state backups, and remove your account, channel, and billing records across our systems. Deletion is handled manually and typically completes within a few business days.

We rely on these third parties to run the service:

• Cloudflare — hosting and storage (Workers, D1, R2) and signup anti-abuse (Turnstile).
• A managed compute partner — hosting for your customer VM.
• Stripe — billing and payments.
• Better Auth on managed Postgres — authentication and account storage.
• Our model gateway and the LLM providers behind your selected model (for example Anthropic, OpenAI, Google, DeepSeek, Moonshot, Zhipu, Alibaba, and xAI — some outside the US) — receive your prompts and conversation content to generate responses. The gateway records usage totals only, not conversation content.
• Resend — verification, password-reset, and invitation emails; receives your email address.
• PostHog (US) — product analytics and in-app support chat; receives account identifiers and your IP address.
• Sentry — error monitoring; receives account identifiers and your IP address.
• Discord, Telegram, Meta (WhatsApp), and Lark operate the messaging platforms themselves and receive your messages directly via their own infrastructure.

Analytics and error data are processed in the United States. We do not send message or conversation content to PostHog or Sentry.

We set a login-session cookie scoped to inference.ai so you stay signed in across Inference AI apps. PostHog sets first-party analytics cookies and local storage on agents.inference.ai for product analytics and in-app support. Analytics and error data are processed in the United States.

We don't sell your data. We don't use your conversations to train models, and we don't authorize our model gateway or providers to. Beyond the sub-processors listed above — which include the model providers that generate your agent's responses and our analytics and error-monitoring tools — we don't share your data with third parties. We don't monitor your messaging channels; our logs record message IDs and sizes, not content.

Questions, deletion requests, or privacy concerns: hello@inference.ai.