Privacy policy

Garuda is operated by Ravan AI. Throughout this policy, we means Ravan AI and Garuda means the product.

When you connect a third-party account through Composio, or add a webhook endpoint, we store the connection record and the endpoint you configured. We do not store the credentials for your third-party accounts; Composio holds those. See sub-processors.

This is what the widget sends when someone opens a chat on a customer's website, and what the server keeps. It is stored for the business that owns the agent.

There are two separate consents, and they do different jobs.

If a visitor agrees to be remembered, the agent can pick up where the last conversation left off. Here is exactly how that works, because the design is the reason it cannot be used for tracking.

• The server generates 32 random bytes. The token is that random value and nothing else — it contains no name, no email address, no device information and no account identifier.
• The widget stores it in the embedding site's own browser storage, under a key that names the specific agent it belongs to. A token from one customer's site is never read by another customer's widget, because the key does not match.
• On the next visit the widget sends that token back. The server does not store the token itself: it derives an identifier using HMAC-SHA256 over a server-held secret, the agent's id and the token, and stores only that.
• Because the agent's id is mixed into the derivation, the same token would produce a completely different identifier for a different agent. There is no shared identifier that could join two customers' records together.

So the token recognises a returning visitor for one agent, on one site. It cannot correlate the same browser across different customers' websites, and we could not build a cross-site profile from it even if we wanted to. That property is covered by a test in our codebase, so a change that broke it would fail the build.

A visitor who wants to be forgotten by an agent can clear that site's browser storage: the token is gone, and the next visit starts as a new one. That does not delete the transcripts already stored for the business — for those, see your rights.

A Garuda agent does not write its answers by itself. To produce a reply, the server sends a request to Google's Gemini API through its OpenAI-compatible endpoint. If you are deciding whether to embed a chat agent on your site, this is the paragraph that matters.

Whatever a visitor types into the chat is part of the conversation and therefore reaches the model provider. Tell your visitors not to type card numbers, passwords or government identifiers into a chat window — no chat agent is the right place for those. Google's handling of the request is governed by Google's own terms for that API.

We do not train any model on your content. Garuda has no training pipeline: your knowledge sources and conversations are used to answer questions in your own workspace and for nothing else.

These are the companies that process data on our behalf so the product can work. We name what each one actually receives rather than listing them as a formality.

Every one of these is a company based in the United States, so assume that data sent to them is processed there. There are no advertising networks and no analytics vendors on this list, because we use none.

If you configure an outbound webhook, Garuda will send lead and conversation events to the URL you chose. Once an event leaves for your endpoint, it is in your hands and your provider's. Our delivery log records the event name, its status and the attempt count — never a visitor's email address or phone number.

Depending on where you live, you may have rights over your personal data — to get a copy of it, to correct it, to have it deleted, or to object to how it is used. We will act on these requests regardless of whether a particular law applies to you. Requests are handled by people, by email, not by an automated flow.

We will acknowledge a request when we read it and tell you what we are doing. We deliberately do not publish a fixed-hours response guarantee, because nothing in the product enforces one and a promise we cannot keep is worse than no promise.

Garuda's own pages set no cookies. There is no cookie banner on this site because there is nothing to consent to: no advertising pixels, no analytics scripts, no third-party trackers. What the product does use is browser storage, and here is all of it.

Widget storage lives under the customer's own website address, not ours, and every key names the single agent it belongs to. If browser storage is unavailable — a private window, or a browser that blocks it — the widget still works; it simply cannot remember anything between visits.

The measures in place today include: passwords stored only as salted PBKDF2-SHA256 hashes; session tokens stored as digests rather than as the tokens themselves and compared in constant time; a widget that only runs on domains the customer has listed; per-address rate limits on every public endpoint; and HTTPS everywhere. The security page goes through each of these in detail.

No service can promise it will never be breached. If a breach affects your data, we will tell you what happened, what we know and what we are doing about it.

Garuda is a business tool and is not directed at children. We do not knowingly collect personal data from a child. If you believe a child's data has ended up in a Garuda workspace, write to info@ravan.ai and we will remove it. Customers must not configure an agent to collect contact details from children.

When this policy changes, the date at the top changes with it. A change that materially affects how we handle personal data will also be announced by email to account holders. Continuing to use Garuda after a change means the updated policy applies to you.

Garuda is operated by Ravan AI. For anything in this policy — a question, a correction, or a request about your data — write to info@ravan.ai. Our terms of service are at garuda.ravan.ai/terms.