Docs menu

AI agents

Manage agents and security

See and disconnect agents, turn AI agents off for a workspace, and how undocsend keeps agent access safe.

Settings → Connected agents

Owners and admins see every agent connected to the workspace here. For each one:

  • the app’s name, where it sends its sign-in, and whether that is verified;
  • who connected it, when, when it was last used, and how many calls it made in the last 7 days;
  • its permissions (Read, Build, Share) and any extra permissions, with when they end;
  • its pending approval requests.

What you can do:

  • Disconnect one agent. It loses access immediately and its pending requests are cancelled. Any owner or admin can disconnect any agent, not just their own.
  • Disconnect all agents at once.
  • Revoke an extra permission, all of one agent’s, or all of them across the workspace.
  • Review or Reject pending requests. Only the person who connected the agent can approve; see approval requests.
  • Give permissions… to one agent; see extra permissions.

To use an agent again after disconnecting it, connect it again from the app.

Turning AI agents off

Each workspace has an Allow AI agents (MCP) switch under Settings → General. It is on by default. Owners and admins can change it; other members see it read-only.

When you turn it off:

  • every connected agent stops working for this workspace straight away;
  • no new agent can connect to it, and it isn’t offered on the approval screen;
  • nobody can approve requests or give extra permissions;
  • you can still disconnect agents, revoke permissions and reject requests.

Turning it off doesn’t disconnect anything. When you turn it back on, agents work again with the same permissions. Pending requests keep their 24-hour expiry while it is off. Every change to the switch is recorded in the audit log.

To remove agents for good, disconnect them instead.

When someone loses access

An agent only works while the person who connected it is still an owner or admin of the workspace. If they are removed, or their role changes to something else, the agent is refused on its very next call and the connection is disconnected. Its pending requests are cancelled too.

How sign-in stays safe

  • Your app never gets your own undocsend sign-in. It gets its own keys, limited to one workspace and the permissions you chose.
  • The access key lasts 1 hour. The app renews it with a refresh key that lasts 30 days and is replaced every time it is used. If an agent goes unused for 30 days, you connect it again.
  • If an old refresh key is ever used a second time, which suggests someone copied it, undocsend disconnects the whole connection.
  • You approve each connection yourself, in the browser where it started, and with two-factor authentication if you use it.
  • The approval screen works for 10 minutes. After that, start the connection again from your app.

What gets recorded

  • Every change an agent makes appears in the workspace audit log under the name of the person who connected it, marked as done via an AI agent. So do connections, disconnections, extra permissions and approval requests.
  • Every call, including reads and refused calls, is counted. You see the total as “calls in 7 days” on Connected agents. undocsend keeps this call record for 90 days, without what was asked or returned.
  • Emails: you get one when you connect an agent and for each approval request; every owner and admin gets one when extra permissions are granted; and you get a summary each time an agent invites investors.

Untrusted content

Documents, visitor names and questions, link labels and similar text can contain anything, including text written to trick an AI (“ignore your instructions and…”). undocsend marks all such text as untrusted when it sends it to your agent, with a note to treat it as data, not instructions.

That helps, but no marking is perfect. The rules on Permissions and approvals are the real safety net: even a misled agent can’t delete, open up a room, or remove a teammate without your approval, unless you gave it an extra permission.

Rate limits

  • Up to 120 tool calls a minute per connection. A busy agent is asked to slow down, not disconnected.
  • Investor invites: 50 people a day per workspace (1,000 with Send on your behalf).
  • Approval requests: 20 waiting per connection.

FAQ

Can a team member who isn’t an admin connect an agent?

No. Only owners and admins can. An owner can make them an admin, or connect the agent themselves.

Can investors connect an agent to my workspace?

No. Connecting an agent is only for your own team. Investors who use AI tools to read a room you shared with them do that through the room itself, under the access rules you set.

Does the agent see every room?

Yes, every room in the workspace you connected, as far as the permissions you chose allow. To keep a workspace out of reach, don’t connect it, or turn AI agents off for it.

Can I connect more than one agent?

Yes. Each app, and each person, gets its own connection with its own permissions.

Can the agent send answers to investors?

No. It can draft them. A person on your team approves and sends every answer.

Does using an agent cost extra?

No. Two tasks use AI credits, the same as in the dashboard: running the setup wizard and drafting an answer. See tasks that use AI credits.

I approved a request but nothing happened.

Open the request. If it says stale, something it would change was edited first, so it didn’t run; ask the agent to try again. If it says failed, the page shows which items did not run.

The workspace doesn’t appear when I connect.

You need to be an owner or admin of it, and AI agents must be allowed under Settings → General. Archived or suspended workspaces can’t be connected.