AI first
AI agents work in Netzo like one more person, with their own credential, the same permissions and a person who decides the changes that matter.
Netzo is built so that an AI agent can work in it safely. An agent does not use a back door: it calls the same actions as the app, with the permissions of its credential, and every change is recorded in its name.
Three ways to work with an agent
| Way | Where the agent works | Acts as | Read |
|---|---|---|---|
| In the tab | The browser agent, on the open app | The signed-in person. The person confirms every write. | WebMCP |
| Connected | Claude, ChatGPT, Cursor or your own agent | Its API key or the person who connected OAuth | MCP |
| In a terminal | An agent with a shell, a script or CI | Its API key or the signed-in person | CLI |
AI proposes, a person decides
An agent never gets more than its credential allows. The actions that need a person wait for one:
| Control | What it does | Read |
|---|---|---|
| Permissions | The role of the credential decides which tables and actions the agent sees. What it cannot do, it does not see. | Access |
can | The agent asks before it acts and gets the permission that decides. | Access |
| Dry run | Every write can run without saving: it runs the rules and the calculations and returns what would change. | Surfaces |
| Approvals | An action with a policy waits for an approver, no matter who asked for it. | Access |
| Activity log | Every change records the credential that made it. | Events |
What an agent knows
An agent does not guess field names. Netzo gives it what it needs to act well:
- The description of each table, with
describe_table: fields, relations, allowed actions and what they mean. - Errors that say how to fix them: a refusal names the rule and the field, the same for a person and an agent.
- Skills: protocols for a complete job, such as to reconcile a bank statement or load a spreadsheet. See Skills.
- MCP apps: tables, charts and cards that the assistant shows in the conversation. See Tools.