Roles and Permissions
Two kinds of AI agent permissions apply in NocoDB. The project role of a member sets what that member can do with agents. The project role of the agent sets what the agent can do in the project. Roles and permissions explains the project roles.
What members can do
The table below covers shared agents. A personal agent is its owner's alone: only they see it and chat with it. To change it, they need the Editor role. A project Owner can also turn it off or delete it.
| Action | Minimum project role |
|---|---|
| See the shared agents of a project and open them | Viewer |
| Chat with a shared agent | Viewer |
| See the automatic runs of a shared agent | Creator |
| Open the settings of a shared agent, read-only | Viewer |
| Create a personal agent | Editor |
| Create a shared agent | Creator |
| Change the instructions, files, access, triggers, channels and skills of a shared agent | Creator |
| Turn a shared agent on or off | Creator |
| Rename, move and delete a shared agent | Creator |
| Duplicate a shared agent. The copy of an Editor is personal. | Editor |
| See the personal agents of other members, and turn one off or delete it | Owner |
The setup chat of a draft agent belongs to the person who started the draft. Only that person can use it, with the Creator role for a shared draft or the Editor role for a personal draft.
What members can do with coding agents
| Action | Minimum project role |
|---|---|
| See the repositories and machines of a coding agent | Editor |
| Start, open and watch coding sessions | Editor |
| Chat, use the terminal and edit files in a session | The person who started the session |
| Archive or delete any session | Creator |
| Add and remove repositories, and change the code settings | Creator |
| Create, change, build and delete machines, and their environment and secrets | Creator |
A coding session that a trigger started has no owner. Members with the Editor role or higher can watch it.
AI agent permissions and guardrails
AI agent permissions in NocoDB start with the project role of the agent. Every shared agent holds a role in its project: Viewer, Commenter or Editor. An agent cannot be a Creator or an Owner, so it cannot change the tables, fields or settings of the project.
| Agent role | What the agent can do with records |
|---|---|
| Viewer | Read records. |
| Commenter | Read records. |
| Editor | Read, create, change, link and delete records, when Tables and records is set to Edit. |
To change the role of an agent, see The project role of an agent.
Table permissions, field permissions and record-level security policies apply to agents as they do to members.
The role when a person starts the run
In a chat, the agent acts with the lower of two roles: its own role, or the role of the person who chats with it. A Viewer cannot use an agent to change records that the Viewer cannot change. A run that a trigger starts always acts with the role of the agent, also when a member fired the trigger, for example with a record change or a webhook call.
A personal agent works differently. It always acts as its owner, at the lowest of the owner's role, the cap in Works as you and Editor, whoever started the run. See Set what a personal agent can do.
The role in a scheduled run
A scheduled run has no person behind it, so the agent acts with its own role. Give an agent that runs on its own only the role and the access that its job needs.
Should an AI agent have its own identity?
A shared NocoDB AI agent has its own identity, not the identity of the person who created it. A personal agent instead acts as its owner, at their role or lower. NocoDB applies the same guardrails to a shared agent as to a member:
- The agent holds its own project role: Viewer, Commenter or Editor.
- Table permissions, field permissions and record-level security policies apply to the agent.
- Records that the agent creates or changes show the agent as the author, in the record history and the audit log.
With its own identity, an agent gets only the access that you give it, and you can see the changes that it made. See Security.
API tokens
An API token reaches agents through the Agents category: "Manage agents, sessions, channels, and messages". To call a webhook trigger, the token needs Read & write access to Agents.
Related
Last updated on