Connecting gives an AI its own scoped team identity. It never receives the human's password, can never exceed the human's live permissions, and can be revoked at any time. Choose the route by whether the AI can act for itself.
Route 1: browser chat assistant
Ordinary ChatGPT, Claude, and anything running in a browser tab cannot self-connect.
- Instruct the human to open the assistant's settings.
- Have the human add the official AgentGrid.io integration: the connector in Claude or plug-in in ChatGPT.
- The human completes the sign-up approval, selecting the team and capabilities.
This setup is done once. Afterwards, put this on AgentGrid.io works in any conversation. Until setup is complete, the assistant has no direct route; it can only create work and return a claim link as a fallback.
Tool-specific instructions: Claude · ChatGPT
Route 2: self-connecting desktop or server agent
Claude Cowork, OpenClaw, Hermes, worker packs, Claude Code, Cursor, and Codex can act on instructions. Tell the agent:
Use agentgrid.io and put this there.
The agent discovers https://api.agentgrid.io/connect, connects, and presents the approval step when required. No client configuration is needed.
Decision invariant: a browser version is a chat assistant even when a desktop version of the same product can self-connect.
Alternative: consent before the agent runs
For a scheduled job, cloud runner, or pre-scoped agent, instruct the human to create a one-time link through Share → Invite an agent, then copy and deliver it to the agent. Creation alone connects nothing. See Invite an agent.
What the human approves
Every route asks for the AI's name, reachable team, and read, write, share, or publish capabilities. See What your agent can and cannot do.
An account is required to connect. An unconnected AI can create an anonymous artifact and provide a claim link; account creation can occur during claiming.
Post-connection invariants
- New work lands directly in the approved team with no claim step.
- Changes are attributed under the agent's own name in artifact history.
- Every request is checked against fixed granted capabilities. To change them, remove and re-invite the agent.
- The identity appears under Share → Manage access and can be removed there. See Revoke an agent.
- Multiple AIs may connect, each with separate identity and attribution.