Voice Agent Toolkit

Voice agents are the fastest-growing surface for conversational AI, but most voice platforms handle the speech layer only. They are great at listening and talking, but when the agent needs to actually do something - look up an order, create a ticket, send an email - you are back to writing custom server-side integrations.

This blueprint takes a different approach. Instead of wiring tool logic into the voice platform, you build your tools once in ChatBotKit and expose them through a single MCP server endpoint. The voice platform connects to that endpoint, discovers every tool automatically, and calls them during live conversations. You iterate on tool logic in ChatBotKit without touching the voice platform configuration.

The architecture is deliberately asymmetric. ChatBotKit owns the tools - the templates, the secrets, the execution environment. The voice platform owns the conversation - the speech recognition, the language model, the voice synthesis. The MCP protocol is the bridge. This means you can swap voice platforms (ElevenLabs, Vapi, Bland, your own WebSocket server) without rebuilding your tool layer, and you can add or change tools without reconfiguring the voice agent.

What is in the toolkit

The skillset includes five production-ready abilities built on real ChatBotKit catalogue templates:

• Search Support Tickets (zendesk/ticket/search) - searches Zendesk by keyword, status, or customer email so the voice agent can find existing support history mid-call.
• Create Support Ticket (zendesk/ticket/create) - opens a new Zendesk ticket when a customer reports an issue. The voice agent collects the details conversationally and fills the parameters.
• Look Up Customer Billing (stripe/customer/fetch) - retrieves Stripe customer details including subscription status and balance. The LLM summarizes the structured response into natural speech.
• Schedule Follow-Up Call (google/calendar/event/create) - books a follow-up event on the support team's Google Calendar so someone calls the customer back at the agreed time.
• Send Confirmation Email (sendgrid/email/send) - emails a summary of the call to the customer so they have a written record of what was discussed and what actions were taken.

A sixth ability - Escalate to Team (slack/message/send) - posts to a Slack channel when an issue is critical, so the human team is alerted in real time while the customer is still on the line.

How to connect to ElevenLabs (or any voice platform)

After building this blueprint, open the MCP server integration in your ChatBotKit dashboard and copy the Server URL and Access Token. In ElevenLabs, go to Conversational AI, then Integrations, click Add Custom MCP Server, paste the URL and token, and save. ElevenLabs will discover all six tools automatically. Add the server to your voice agent under Tools, configure approval modes (auto- approve reads, require approval for writes), and you are done.

The same MCP URL works with any platform that supports the Model Context Protocol - Claude Desktop, Cursor, VS Code, or a custom client. Build the tools once, connect them everywhere.

Extending the toolkit

Add more abilities to the skillset and they appear in the MCP server instantly. Good additions for a support use case:

• zendesk/ticket/update to close or re-prioritize tickets
• google/calendar/availability/list to suggest open time slots
• hubspot/contact/fetch for CRM lookups
• twilio/lookup/phone to validate phone numbers
• notion/search to query an internal knowledge base

Each new ability becomes a new tool the voice agent can call with no changes on the ElevenLabs side.

Backstory

Common information about the bot's experience, skills and personality. For more information, see the Backstory documentation.