Build a voice agent
Pick an agent on the left, or create a new one to start.
How to build (the workflow)
- Info tab: name your agent and pick a voice.
- Variables tab: declare every value your APIs and prompt will use (e.g.
employeeId,companyId). Mark where each comes from (JWT claim, computed, tool response). - APIs tab: add one entry per HRMS endpoint. Use
{{variableName}}placeholders in URL/headers/body. Map response fields to variables. - Prompt tab: paste your system prompt. Reference variables with
{{employee_name}}, etc. - Embed tab: click Activate, then copy the embed snippet into the host page.
Info
Variables
APIs
Prompt
Embed
LLM (brain)
Groq Llama is good for high-volume happy-path conversations but sometimes skips protocol steps, hallucinates dates, or emits malformed tool calls (we have client-side patches for these).
GPT-4o / Claude Sonnet follow multi-step protocols much more reliably — recommended for production HR / finance use cases.
Requires the right key in
GPT-4o / Claude Sonnet follow multi-step protocols much more reliably — recommended for production HR / finance use cases.
Requires the right key in
.env: OPENAI_API_KEY for GPT, ANTHROPIC_API_KEY for Claude, GROQ_API_KEY for Llama.
Voice
Browse voices at elevenlabs.io/app/voice-library. Paste the voice id here.
Declare values your APIs and prompt reference. Use {{employeeId}} syntax in API URLs, headers, body, and prompt text.
Each entry becomes a tool the agent can call. Use {{variableName}} anywhere a value is needed (URL, header value, body, etc.).
System prompt (markdown OK). The agent reads this every turn. Reference variables and API names directly.
1. Click Activate below — this makes this agent the live one.
2. Copy the embed snippet into the host page (e.g. HRopal portal dashboard).
3. Replace USER_JWT_FROM_HRopal_PORTAL with the actual logged-in JWT at runtime.