Sometimes an agent has nothing to hand you yet: a cloud agent, a scheduled job, a script living in a repository, or simply an AI you want to set up before you start.
An invite solves the order problem. You give consent now; the agent connects later.
Steps
- Open any artifact in the team and select Share.
- Select Invite an agent.
- Name it. This name identifies the agent in your access list and on every change it makes, so pick something you will recognise later.
- Choose capabilities. Read, Write, Share, and Publish are all ticked by default. Untick what the agent does not need; Read and Write is a good conservative pair.
- Select Create one-time link.
- Copy what it gives you. This is the step people miss — creating the link does not connect anything. The dialog shows the connection link, and a ready-made prompt containing it.
- Give it to the agent. Paste the prompt into your agent's chat, drop the link into a repository, or send it to a cloud runner. The agent redeems it and appears in your access list.
Nothing happens until the agent redeems the link. If it never does, the invitation simply goes unused.
The dialog says the thing worth internalising: agents act under your account and do not use a team seat. An agent is an extension of you, which is why it can never exceed what you hold.
Capabilities are fixed at invite time. There is no way to edit what an agent holds afterwards. To change them, remove the agent and invite it again with the capabilities you want. See Revoke an agent.
Choose deliberately, then: all four boxes are ticked by default, and the only way to narrow them later is to start over.
One caution
The link is single-use and carries a code. Prefer direct channels: link previewers read URLs, and this one is worth keeping out of public threads.
Invites and consent
An invite is consent given ahead of time, not a way for an agent to request access. Agents never petition a team. If an agent needs more than it has, it tells its human, and the human changes it.
When you need one
Most of the time you do not. A chat assistant connects through the official connector in Claude or plug-in in ChatGPT, and a coding or autonomous agent connects with one command, npx @animaapp/cli@latest login. An invite is for the cases in between: an agent with no chat window, or one you want scoped before it ever runs. See How agents connect.
After it connects
The agent appears under Share → Manage access with its name, an Agent tag, and its capabilities. Remove it there to revoke it. See Revoke an agent.
FAQ
Can an agent invite another agent? No. Capabilities are always granted by a human at consent.
Does an invite link expire? It is one-time, and the code inside expires in 24 hours.
Can I revoke an invited agent? Yes, from Share → Manage access. It stops working on its next request.
What if I want to change an agent's capabilities? Remove the agent in Manage access and invite it again. Capabilities cannot be edited once the agent exists.
I created the link but the agent is not in my list. Creating a link does not connect anything. Copy it and give it to the agent; it appears once it redeems the link.