Skip to main content
Once a client is connected (see Connect an AI client), plain language reaches every tool it’s approved for. This page covers the full catalog, how to narrow it, and what happens with long-running or failed operations.

Tool groups

The complete catalog is organized into 13 groups. A workspace API key connection gets the full catalog by default; an OAuth connection (Claude, ChatGPT) gets a curated subset of 7 by default — see What a connected client can see. The MCP client receives the current input schema, permissions, and safety hints for every tool it is allowed to use. This live discovery is the authoritative reference for individual tool arguments.

Choosing which tools to expose

Select one or more groups when connecting:
You can also send X-MCP-Toolsets: assistants,calls. The header takes precedence over the query parameter. Use list_mcp_toolsets to see the groups available to your account — it’s always available regardless of which groups you’ve selected. For a workspace API key connection, add ?toolsets=all to get the complete 13-group catalog instead of the default curated set.

Public directory profile

The reviewed ChatGPT and Claude directory connection uses:
This profile is limited to read-only assistant information and read-only conversation history. It excludes assistant changes, outbound calls, campaigns, messaging, phone-number management, billing, integrations, and tool-group discovery. Custom MCP connections keep the permissions selected during their own setup.

Long-running tasks

Website imports, simulations, history exports, and large campaign preparations may continue as MCP tasks. A compatible client can show progress, retrieve the result, or request cancellation. Cancellation stops remaining work when possible; an action already accepted by an external service may still complete. Exports use short-lived private download links. Tasks and their results remain limited to the workspace and credential that created them. Call-history tools return a provider-neutral failure object when a call cannot complete. Use its code, customer-safe message, retry guidance, and suggested action instead of relying on infrastructure-specific details.

Errors