Security
Each run stays in its own sandbox and acts only as the person who started it. Writes wait for your approval, and the audit log records every action.
01 · Security
Secure by construction.
Evergreen is built around what an agent can’t do, so you don’t have to trust it to behave.
You signed in
Single sign-on
Sign in with SSO and device checks, and Evergreen checks again every hour.
Sandbox · deny-by-default network
Agent works in its sandbox
The sandbox is the fence
Each run gets its own sandbox, quota, and deny-by-default network.
Run token short-lived, this run only
No standing credentials
Each run gets short-lived tokens for that run only, fetched one call at a time.
Approval before any write
Writes wait for you
The agent reads while it plans. Writes wait for your approval.
Honest authorship
The platform sets commit identity. The agent can’t pose as anyone.
Audit log append-only
An audit trail
Every action, who took it, and their role, in an append-only log.
02 · FAQ
Questions people ask.
Do you or the model providers train on our code?
Carta won’t use your code, prompts, or other content to train or fine-tune AI models without your express permission. When Evergreen runs a model on Carta’s account, our agreements or settings with that model provider stop it from training its models on personal data in your content. On your own key, your agreement with your provider decides what it does with your data. Our Terms and DPA set out these commitments.
Do we pay for everyone on the team?
No. There are no seat fees, so invite everyone, whether they start sessions or only watch them. You pay only for the tokens, agent time, and workspace storage your sessions use, for the embeddings and reranking behind your workspace’s search and knowledge, and for the model calls Evergreen makes for your workspace: titles, summaries, search intent, and reviews.
Is there a free trial?
A workspace you own pays no platform fee, agent time, workspace storage, or search for seven days, counted from its first session. On your own Anthropic, OpenAI, or xAI key you don’t need a card, and the week starts on its own when you run your first session. Models and knowledge are still billed from the first turn: models at the provider’s price (or by your provider, on your own key), and knowledge at AWS’s price. On Evergreen’s models, that means you’ll add a card and funds before your first session. Each person gets one free week in the first workspace they own that runs a session.
Do you mark up the models?
Never. Without your own key, you pay exactly what the model provider charges us, and when we get a discount, so do you. Evergreen’s own price is 15¢ per million tokens, with agent time and workspace storage beside it.
Can I use my Claude Pro, Max, or Team plan with Evergreen?
A Max or Team plan, yes: both include monthly Claude API credits. Link a Claude Console organization to your plan to claim them and create an API key there; a workspace owner or admin then sets it as the workspace’s own Anthropic key. Every Claude run in the workspace, not only yours, then spends those credits first, and Anthropic bills that organization directly for anything past them. A Pro plan includes no API credits, and signing in with a plan isn’t supported. Evergreen’s own fee, agent time, workspace storage, search and knowledge, and the model calls Evergreen makes for your workspace that your key does not serve stay on your Evergreen bill.
Does a cheaper model make Evergreen cheaper?
The model’s part of the bill, yes: models pass through at the provider’s price, so a cheaper model costs less. Evergreen’s own fee is 15¢ per million tokens on every model, and agent time doesn’t depend on the model either.
What does workspace storage cost?
35¢ per GB-month, only while a session’s workspace is kept. Archiving a session or marking it done frees its workspace straight away; otherwise it is removed on its own within 24 to 72 hours.
Can I cap what we spend?
Yes. Set a daily or monthly limit for the whole workspace, for everyone, for an audience, or for one person. You choose what happens at the limit: stop sessions, stop new sessions only, or only send a notice. Someone who reaches a limit can ask an admin for more.