elevenlabs-agents_skill

Overview

This skill helps you build, configure, and troubleshoot conversational AI voice agents using the ElevenLabs Agents Platform. It covers SDK selection, agent prompts, RAG knowledge bases, tools (client/server/MCP/system), versioning and A/B testing, and security patterns like MCP approval modes and secret handling. It also documents common breaking changes, package migrations, and 34 documented errors to prevent.

How this skill works

The skill inspects agent configuration, SDK usage, and runtime integration points to validate correct parameter names, tool formats, and dynamic variable usage. It guides selection of the correct ElevenLabs packages per environment, ensures TTS/ASR/turn-taking settings are valid, and surfaces known gotchas (deprecated fields, removed models, browser-only issues). It also outlines workflows for RAG indexing, voice management, and tool execution paths (client, server, MCP, system).

When to use it

• Building production voice agents, phone systems, or multi-voice conversational UIs.
• Setting up agent versioning, A/B testing, subagents, or branching workflows.
• Configuring RAG knowledge bases, document indexing, and retrieval settings.
• Securing MCP integrations, service-account keys, and secret variables.
• Troubleshooting SDK errors: deprecated fields, CSP/webhook issues, or 'child_process' browser errors.

Best practices

• Use the correct ElevenLabs package for your environment (elevenlabs-js server only; @elevenlabs/react for React; react-native for mobile).
• Always use camelCase parameters in JS/TypeScript SDKs (modelId, voiceId, outputFormat).
• Migrate away from legacy prompt.tools to tool_ids and built_in_tools; never send both formats.
• Provide all dynamic variables referenced in prompts and mark secrets for header-only use.
• Use Turbo v2/v2.5 for TTS (v1 removed) and test voice switches to measure added latency.

Example use cases

• Deploy a multilingual customer-support voice agent with RAG-enabled policy documents and live tool calls to CRM APIs.
• Create an AI phone companion that uses turn-taking 'patient' mode for intake and 'eager' for quick responses.
• Set up A/B agent versions to test different personalities, prompts, or voices using agents-cli for deployment.
• Implement a proxy server for browser apps to call elevenlabs-js safely and avoid child_process bundling errors.
• Integrate MCP servers with fine-grained approval for database queries and track tool_latency_secs for SLAs.