Actions let an agent do more than speak. Add only the tools required for the workflow, give each one a narrow responsibility, and test its failure behavior before publishing.
Built-in actions
Define the contract
Give the function a lowercase snake_case name and explain exactly when the agent should use it.
Configure the request
Choose the method and URL, then add headers, query parameters, and a JSON object schema for inputs.
Handle the response
Decide which fields should be stored as variables and what the agent should say before, during, or after execution. You can play a typing sound while it runs.
Test errors
Verify timeout, non-success response, missing field, and malformed-response behavior.
Do not place credentials directly in prompts or tool configuration. Store encrypted values under Settings → Secrets and reference them through the supported configuration.
- A cold transfer connects the destination without an agent debrief.
- A warm transfer lets the AI provide a debrief to the next agent.
Choose SIP invite or SIP refer according to your provider support. For SIP invite, configure the displayed caller ID and transfer ring duration. Add custom SIP headers only when the receiving system requires them.
Set the expected digit count, termination key, prompt, timeout, confirmation behavior, and variable that stores the result. Test slow entry, early termination, invalid digits, and timeout.
MCP servers
An agent can connect to an MCP server and expose selected MCP tools. Configure the server URL, headers, query parameters, and timeout, then choose the tools the agent may call. A workspace-global MCP can be available across agents; an agent-specific MCP stays scoped to that agent.
Treat an MCP server like any other privileged integration: grant the smallest useful tool set and verify its authentication and failure behavior.
Calendar connections
Calendar actions require a connected Cal.com or Calendly account. Manage connections under Integrations before adding calendar actions to an agent. 
