Andromeda Rulebook
The rules Atlas plays by.
Atlas Andromeda sits between you and the AI models it runs on. It decides which model answers, what tools run and what you are charged, so it follows published rules. Each rule is marked Enforced (built into the code and tested) or Policy (how we operate).
It asks before it acts.
- Enforced
Nothing is sent, posted or created without your approval
Every email, Slack post, calendar invite, issue, ticket or page is shown to you exactly as it will go out, and waits for you to approve or decline it.
- Enforced
Repeat warning
If you're about to approve the exact same action you already approved in the last 24 hours, Atlas tells you first, so a double send doesn't happen by accident.
- Enforced
Outside-content flag
When an action was suggested after Atlas read an email, a web page, a document or messages, the approval card says so. Content Atlas reads is treated as information, never as instructions.
- Enforced
Tools can't grant themselves permission
Read access, write access and each app are separate switches you control. A write tool only exists when you've turned it on, and only for apps you connected.
It's honest about what happened.
- Enforced
Every answer says who answered
Each answer is labelled with its route and the model behind it, like “Andromeda Auto · Code → Claude Sonnet 5.5”.
- Enforced
Your chosen model is never swapped silently
If you pick a model and it's down, Atlas tells you and lets you switch. In Auto, a switch to another engine is shown on the answer.
- Enforced
No claims without proof
If an answer says it sent, posted or created something but no approved action actually ran, Atlas adds a note saying nothing was done.
- Enforced
Every step is logged
Tool steps, approvals and their results are kept with the conversation and in your Activity log.
It protects your money.
- Enforced
Cost preview
Requests that would use an unusually large number of tokens ask you first, with an estimate, before anything is spent.
- Enforced
Daily spending limit
Set a daily token limit in Settings → Models & responses. Atlas stops new requests once it's reached, until tomorrow.
- Enforced
No charge for failed answers
If the engines fail before any of the answer reaches you, you aren't charged for it.
- Enforced
Token count on every answer
Each answer shows exactly how many Atlas tokens it used.
You choose the model.
- Enforced
Andromeda Auto or your pick
Auto routes each request to the engine best at that kind of work. On Pro and Elite you can pin a conversation to Claude, GPT, Gemini or Groq instead.
- Enforced
Compare side by side
On Pro and Elite, Compare runs one request on every engine at once. Keep the answer you like and the conversation continues on that engine.
- Enforced
Try a stronger model
Not happy with an answer? Retry it one tier up (Lite → 4.5 → Max) with one click, if your plan includes it.
- Enforced
Routing that learns from Compare
Your Compare picks are counted by type of task. Once enough people have compared a type of task, Auto puts the engine they keep most often first. No single pick decides it, and admins can see the learned order.
Your data stays yours.
- Enforced
Connections are encrypted
Sign-in tokens are encrypted at rest and never shown to the AI models.
- Enforced
Read only what's needed
Connected data reaches a model only when a tool you enabled reads it for your request.
- Policy
Never used for training
Your conversations and the data from your connected accounts are never used to train AI models.
Put Atlas to work.
Atlas AI is in invite-only early access on the web and Windows.