How it works
One assistant. Many models. You still know where your data went.
Most AI tools tie you to a single model. Friday doesn't. He reads what you asked for and sends it to whichever model suits it — then tells you which one answered. This page explains how that choice is made, and exactly which companies are involved.
How a message gets answered
- 1Friday reads the request. A quick question and a long piece of reasoning need different models. He decides which this is before spending anything.
- 2He checks what the request touches. If it involves your email, calendar, notes or saved memories, that changes which providers are allowed to see it — see below.
- 3He picks a model and answers. Simple work goes to a fast, inexpensive model. Hard work goes to a frontier one. The point is to stop spending frontier prices on questions that don't need them.
- 4If a provider is down, he keeps working. Rather than failing, Friday falls through to another model that is approved for whatever the request contains.
What happens to private things
Some of what you give Friday is more sensitive than the rest — your inbox, your calendar, your notes, the documents you upload. When a request involves those, Friday restricts it to providers we've approved for it. Providers we haven't approved are not sent the request, and are not offered the tools that could reach it.
When you give a team of agents a job to work through, that path is stricter still: it only uses approved providers, whether or not the content looks sensitive, and if none are available the job fails and tells you rather than quietly using someone else. We are still extending that same restriction to the remaining background jobs, so we will not yet claim it holds everywhere — our Privacy Policy says the same, and we would rather the two pages agree than sound impressive.
Who sees your data
Every company below receives something when you use Friday — a message, a file, a search term, or just the fact that you loaded a page. They are all here, including the ones whose terms we don't like.
| Company | Role | Based in | Trains on your data? | Sees private content? |
|---|---|---|---|---|
| Anthropic (Claude) | Primary | United States | No | Yes |
| DeepSeek | Fallback / routed | China (PRC) | May train | No |
| Google (Gemini) | In the catalogue — not wired to anything | United States | Unverified | No |
| Moonshot (Kimi) | In the catalogue — not wired to anything | China (PRC) | Unverified | No |
| OpenAI (GPT) | Fallback / routed | United States | No | Yes |
| Alibaba (Qwen) | In the catalogue — not wired to anything | China (PRC) | Unverified | No |
| Z.ai / Zhipu (GLM) | Fallback / routed | China (PRC) | Unverified | No |
| Browserbase | Supporting — when used | United States | Unverified | Yes |
| Cloudflare R2 | Supporting — when used | United States | Unverified | Yes |
| Firecrawl | Supporting — when used | United States | Unverified | Yes |
| Google Fonts | Supporting — always | United States | Unverified | No |
| jsDelivr | Supporting — always | United States | Unverified | No |
| Perplexity | Supporting — when used | United States | No | Yes |
| Render | Supporting — always | United States | Unverified | Yes |
| Stripe | Supporting — opt-in | United States | Unverified | No |
| Twilio | Supporting — opt-in | United States | Unverified | Yes |
| Voyage AI | Supporting — when used | United States | May train | Yes |
Accounts you connect
The table above is companies we chose. Connecting your own accounts adds more, and you should know what that does before you do it.
If you connect Gmail, Google Calendar or Notion, Friday reads from them on your behalf. That doesn't hand your mail to a company that didn't already have it — Google already holds your Gmail. What it does mean is that anything Friday reads from those accounts can become part of a message to a model, and is then subject to everything above: it is treated as private, so only approved providers can see it.
Friday can also connect to other tools through a standard called MCP. Those appear in a directory in the app, and none of them receive anything until you connect one yourself. When you do, you are choosing to add that company to this list. Disconnecting stops it.
What we don't claim
We can't encrypt your prompt from the model that answers it. No provider can. Any model that reads your message receives it in readable form. What we control is which companies that can be.
"We don't train on your data" is about us and our providers, not everyone. We never train on your content. Our main providers don't either, under their business terms. Some fallback providers may — which is why they're in the table, and why private content doesn't go to them.
Providers keep data for a while. Our main providers retain content for a limited period for abuse monitoring. Reducing that means a zero-retention agreement with each of them, and each has a cost — one of them would mean giving up the strongest model we offer. We have not made those trades yet, so we are not going to imply they are in progress, and we will not tell you data is never stored.
Friday is in free beta. Things change, and we'll tell you when they do. You can export or delete everything at any time.
Questions
If you want to know which provider handled a particular request, or you'd rather we switched a provider off for your account entirely, email youssef@broomit.net and we'll do it.