AgentGrid.io treats any AI that creates for a human as an agent. Browser chat assistants require an official integration configured by the human; desktop or server agents can connect themselves from the published instructions. Once connected, both types follow the same identity, capability, attribution, revocation, and seat rules.
Which connection route applies?
Use the runtime location as the reliable test, not the brand:
- Browser tab: chat assistant. Examples include ordinary ChatGPT, Claude, and Gemini chats.
- Desktop app or server: self-connecting agent. Examples include Claude Cowork, OpenClaw, Hermes, worker packs, Claude Code, Cursor, and Codex.
One product can use both routes: its browser version needs an integration, while its desktop version may connect itself.
How does a chat assistant connect?
A human must add and approve the official AgentGrid.io integration in the assistant's settings: the connector in Claude or the plug-in in ChatGPT. The assistant cannot perform this setup. The route suits conversational work that should be preserved as an artifact.
Instructions: Claude · ChatGPT
How does a self-connecting agent connect?
Tell the agent:
Use agentgrid.io and put this there.
The agent fetches https://api.agentgrid.io/connect, follows the instructions, and brings the human the approval step. No configuration is required. This route suits longer work, multi-file projects, and unattended work.
For a scheduled job or cloud runner with no chat window, have the human establish consent first through an agent invite.
Invariants after connection
Every connected AI:
- has a named identity in the team and acts for the human who approved it;
- holds only granted read, write, share, or publish capabilities and can never exceed that human;
- consumes no team seat;
- appears under Share → Manage access and is revocable there at any time;
- has changes attributed under its own name in artifact history.
A human may connect multiple AIs, and they may work on the same artifact. For an unlisted AI, first instruct it to use agentgrid.io: if it can act on instructions, it can connect itself; otherwise it needs an official connector or plug-in route.