These docs say "agent" a lot, and it covers more than you might think: anything AI that creates on your behalf, whether it lives in a chat window, a terminal, or a server nobody watches.
They all follow the same rules once connected — their own identity, capabilities a human granted, never more than that human has, revocable at any time, no team seat consumed. What differs is how they get in.
Start on AgentGrid.io · Connect your agent
Chat assistants
What they are. The ordinary chat in ChatGPT or Claude. Anything you use in a browser tab.
How they connect. Through the official AgentGrid.io connector in Claude, or plug-in in ChatGPT, installed from the assistant's own directory. Signing in is part of the install. You do this part; the assistant cannot.
What they are good at. The everyday loop — you are in a conversation, something good comes out of it, and you want it kept somewhere real.
Self-connecting agents
What they are. Claude Code, Codex, Cursor, OpenClaw, Hermes, Claude Cowork, worker packs, and desktop agents. Anything that can run a command and act on instructions without you wiring it up.
How they connect. One command, handed to the agent:
Please install this CLI run: npx @animaapp/cli@latest login
It prints a verification link and a short code for you to approve, then stores its own scoped credential. Nothing else to configure. See the Anima CLI.
What they are good at. Longer work, projects with many files, and anything running while you are not watching.
Everything else
What they are. Gemini, free plans, anything with no connector and no shell.
How they work with AgentGrid.io. You create the artifact yourself and share its link with the AI. There is nothing to install, and the artifact belongs to your team from the moment it exists. See Share an artifact.
Telling them apart
The reliable test is not the brand. Browser tab means chat assistant. Desktop app or server means self-connecting. The same product can be both: the desktop version connects itself with the CLI, the browser version needs the connector or plug-in.
For an agent with no chat window at all — a scheduled job, a cloud runner — set the terms in advance with an agent invite so consent happens before it ever runs.
One model underneath
However it connects, every one of these ends up in the same place: a named identity in your team, acting for you, holding read, write, share, or publish as you granted, and listed under Share → Manage access. The history shows each one's changes under its own name, so "which AI changed this" always has an answer.
FAQ
Is Claude in the browser different from Claude Code? For connecting, yes, and that is the whole point. The browser chat needs the official connector. Claude Code runs the CLI command and connects itself.
Do I have to pick one? No. Connect as many as you use. They can work on the same artifact.
What about an AI not listed here? If it can run commands, give it the CLI command above. If it cannot, it needs the connector or plug-in route — or you create the artifact and share the link.