Governance & policies
Who can do what, in each workspace, enforced in code
The organization is the company on the platform, isolated from all others; the workspace is a division within it, and it owns documents, agents, keys, budget and policy. People join with a role (owner, admin or member); systems join with an API key, which is valid for a single workspace and does only what its scopes allow.
Workspace policies are versioned: what to do with each sensitive data category before a model call, how many requests per minute and concurrent calls, which tools are blocked. The same policy applies to chat, agents and API calls, and every change goes into the audit log.
How it works
- Roles: the owner manages the whole organization, products and budget; the admin manages the workspace's keys, allowed models and integrations; the member uses the products.
- API keys scoped by function: a key that can only search knowledge cannot call models or run agents.
- Keys valid for 90 days by default (up to 365), rotation with an overlap of up to 24 hours and immediate revocation.
- Sensitive data by category (CPF, CNPJ, card, account, Pix, boleto, secrets, email, phone): block, mask or allow.
- Per-workspace limits: 60 requests per minute and 20 concurrent calls by default, plus tool blocking.
- Separation between organizations in the database, through access rules enforced row by row.
Example use case
A company creates one workspace for the support team and another for finance. In the finance workspace, the admin blocks Pix keys and bank accounts in the sensitive data policy and allows only one model. The billing system gets a key with the scope to run flows, and nothing more. When someone changes the policy, the previous and new versions are kept in the audit log with the name of whoever made the change.
What it does not do
- A key is valid for a single workspace, and each workspace has up to 20 active keys.
- The audit export scope can only be granted by someone who administers the whole organization.
- The policy covers what goes through the platform, the browser extension or the perimeter proxy; AI usage outside them is not seen.
- It does not replace a formal certification.