Skip to main content

​
What is this?

The full catalog of tools the Workspace MCP server can expose to a connected client (Claude, Cursor, VS Code, or any other MCP client), one page per area. This exact list — same names, same descriptions, same order — is always visible live in-app under Developer → Reference → MCP, generated from the same tool registry as this page. A given API key’s actual available tools are narrower: only tools whose required permission the key’s owner holds are ever offered, re-checked on every call, and only inside the one channel the connection was made for.
MCP must be enabled on your installation and included in your plan. Without it the Developer Portal’s MCP tab is hidden or shows an Upgrade badge, and the server answers 403.Only non-destructive tools are ever exposed over MCP — nothing that deletes or cancels something permanently is offered here, no matter what permissions the key has. Every Write tool also needs your MCP client’s own approval step before it runs (the same tool-approval prompt any MCP client shows), on top of your workspace permissions.

​
What can I do here?

  • Look up which tools an assistant gets once connected, before you connect it
  • Check which permission a team member needs for a given tool to appear for their API key
  • See which tools are Write (need your client’s approval before running) versus Read (run immediately)
  • See which tools reach customers (they are labelled Reaches customers in the Developer Portal): sending a template or Direct Send message, creating a campaign, triggering or activating an automation, and publishing an AI agent

​
Tool areas

​
Troubleshooting / Technical Notes

  • A tool I expect isn’t in the list the assistant sees. The MCP server only offers tools whose required permission the API key’s owner holds, re-checked on every call — grant the permission in the “Permission needed” column above (see Team), then reconnect or retry.
  • A tool that changes something (Write) isn’t in this catalog at all. Some write actions — anything that deletes or cancels permanently, such as cancelling an event registration — are never exposed over MCP by design and stay dashboard-only, regardless of permissions.
  • A Write tool call seems to hang or never take effect. Your MCP client (Claude Code, Claude Desktop, Cursor, VS Code) asks you to approve each Write tool call before it runs — check for a pending approval prompt in your client.
  • create_ai_agent, create_automation, or build_ai_agent worked but the agent or automation isn’t answering customers. By design: new AI agents and automations are created inactive (as drafts). Test the agent in the dashboard Playground, run check_agent_readiness, then use publish_ai_agent; or review the flow in the dashboard and use activate_automation. Publishing and activating only happen when you explicitly ask.
  • update_automation is refused. It only replaces the flow of an automation that is off and has no runs in progress. Use pause_automation first, then retry. Automations using steps it can’t describe (payments, catalog, carousels, Google Sheets / Calendar, Shopify / WooCommerce, sequences) are built in the dashboard flow builder.
  • update_agent_settings won’t turn guardrails off. Over MCP, guardrails can only be switched on. Turning them off is a dashboard action, so a tool result or document the assistant reads can never talk it into removing the safety checks.
  • test_ai_agent used AI credits. It runs the real agent, so each test message costs credits like a customer message would. It acts on a labelled test contact only; set endSession on the last turn to remove that contact.
  • A tool I added with add_external_capability is off. If the MCP server refused the login (or needed one and none was given), its tools are saved as needs_credentials and stay off until you add the login under the agent’s External Capabilities in the dashboard. list_external_capabilities shows the status.
  • update_ai_agent replaced instructions I wanted to keep. systemInstructions replaces the whole text. Read the current instructions with get_ai_agent first and send the complete new text.
  • set_handoff_assignment changed other agents. The handoff recipients are a setting of the whole channel, shared by every AI agent on it. Turning handoff on or off for one agent is set_agent_capability with handoff_to_human.
  • direct_send is rejected. Direct Send is a Meta beta and is restricted per WhatsApp Business Account; it only covers utility and authentication messages, never marketing. If a matching approved template exists, use send_template.
  • The list here doesn’t match what I see in-app. It should always match — both come from the same tool registry. If they differ, refresh the Developer Portal’s MCP tab and tell us.
Last modified on October 29, 2026