Where a message goes, and who can see it
Not a slogan, the actual path: our servers, the inference provider, and, when the model searches, a search provider. Nobody else.
Every AI chat sends your words somewhere. Most products describe that vaguely. This page follows one message from the moment you press send to the moment the reply comes back, and names each party that handles it.
How it actually works
Step 1: it reaches our servers
Your message travels over HTTPS to servers we run in Canada, a country the European Commission recognizes as giving EU-level data protection. It is held in memory while the reply is generated. Nothing is written to disk at this point.
Step 2: the request is built in memory
To generate a reply, the server assembles the request in memory, in plain text: your messages, the assistant's earlier replies, the text of any link you pasted, and the instructions for the pipeline you chose. This is the moment the text is readable, and it is why we say encrypted at rest, not end-to-end.
Step 3: it goes to the inference provider
The request is sent over HTTPS to Fireworks AI, which runs the open-weights models. Fireworks states that it keeps prompts and outputs only in memory for the duration of the request, does not write them to persistent storage, and does not use them to train models. That is their published policy, and we chose them for it. We cannot see inside their systems, so we report it as their commitment, not ours.
Step 4, when the model searches: a search provider
With the WebSearch pipeline selected, the model decides per question whether to search. When it does, the search query goes to Brave Search, which returns results and, per its policy, does not keep the query for training. The query may contain words from your message. Every search and fetch shows as a badge on the reply, so you know when this happened. With another pipeline selected, nothing reaches any search provider.
Step 5: the reply comes back, and is stored only if you save
The reply is shown to you. If you have saved the conversation, which requires an account, your message and the reply are written to the database as AES-256 ciphertext. If not, nothing is written: the conversation lives only in your browser until you save it. Fetched pages and search results are discarded after the turn. Deleting a saved conversation removes both sides permanently.
Who does not see it
Stripe sees your name, email, and payment details, never your chats. The email service that sends sign-in links sees your address. Analytics and advertising tools, if you accept their cookies, get page-level usage data and do not receive conversation content. Nobody at Berges reads your conversations, and no model is trained on them.
Who sees your conversations, in plain terms
Is it end-to-end encrypted?
No. End-to-end would mean nobody but you could ever read the message, and then no model could reply to it. Your messages are encrypted at rest and in transit, and decrypted in memory to generate each reply.
Can Fireworks read my messages?
The text passes through their systems in memory to generate the reply. Their policy is that it is not logged or stored and not used for training. We rely on that commitment the same way you rely on ours.
Does anything leave the EU?
Yes. Our servers are in Canada, which the European Commission recognizes as giving EU-level data protection. Fireworks may process the request in other countries, and so may Brave when WebSearch is on. The privacy policy covers this in full.
What if you change providers?
The privacy policy names the current provider, and we update it if that changes. The rule we apply does not change: the provider must not retain your conversations or train on them.