ElevenAgents bills by call minutes rather than shared credits, with the Free plan including 15 minutes of calls per month. The Creator tier costs $22 monthly and includes 275 minutes, while Business runs $990 for 12,375 minutes. Additional minutes cost $0.080 each, and LLM usage bills separately on top of those platform fees. This pricing structure makes high-volume generation expensive if you do not monitor your character and dubbing volume carefully before choosing a tier.
Building the Agent Configuration
You build your first agent in five minutes using the web platform or CLI, requiring no coding. Start by installing the CLI via Homebrew or Scoop, then run elevenlabs auth login to authenticate through OAuth. Initialize a project with the customer-service template, which sets a low temperature of 0.1 for consistent answers and a 30-minute conversation limit. The system prompt acts as the operating manual, so include an honesty clause that tells the agent to use only the knowledge base and admit when information is missing. A voice agent that confabulates pricing sounds authoritative while being wrong, which is worse than silence. Spoken answers past two sentences lose the listener, so keep responses tight.
Upload documents in PDF, Markdown, or HTML formats up to 20MB per file. Small documents under 300,000 characters ride in full context, always available on every turn. Larger files go through RAG, where indexing adds roughly 250ms of latency per answer but keeps large knowledge bases usable. RAG limits vary by workspace tier: 1MB on Free, 2MB on Starter, 20MB on Creator, and 100MB on Pro. Test the agent using the dashboard button, running it through questions users actually ask, including half-formed sentences and slang. Define evaluation criteria like "the assistant was able to answer all queries" so every transcript gets graded automatically against your definition of success.
Deployment and Latency Optimization
Deploy the agent as a web widget by generating an embed snippet from the CLI and pasting two lines of HTML into your page body. Set the Allowlist in the Security tab to your own domains to prevent others from hotlinking your widget and spending your call minutes. For phone integration, connect to Twilio or a SIP trunk later. To minimize latency, use the Turbo v2 model designed for real-time applications. Implement the streaming API to transmit audio as it generates, reducing time to the first byte. Adjust the optimize_streaming_latency parameter to prioritize speed over audio quality if needed. Premade voices generate speech faster than instant voice clones, which have higher latency. Reuse HTTPS sessions to bypass SSL/TLS handshakes and reduce overhead for subsequent requests.
| Plan | Price | Included Calls |
|---|---|---|
| Free | $0/month | 15 minutes |
| Creator | $22/month | 275 minutes |
| Business | $990/month | 12,375 minutes |
The platform co-locates transcription, understanding, generation, and synthesis components to minimize latency. Scribe v2 Realtime transcribes speech in around 150ms across 90+ languages. However, credit-based pricing scales with usage, so heavy TTS or dubbing burns through allowances quickly. If you need finished content rather than just audio, this tool stops at the voice track. Does the 250ms RAG latency impact user satisfaction in fast-paced support scenarios?
