ElevenLabs Agents Platform

Local MCP server (stdio) that gives Claude full control of the ElevenLabs Agents Platform (Conversational AI): agents, conversations, knowledge base, tools, tests, telephony, batch calling, and workspace settings.

Tools (50)

Group	Tools
Agents	el_list_agents, el_get_agent, el_create_agent, el_update_agent, el_delete_agent, el_duplicate_agent, el_get_agent_link, el_simulate_conversation
Conversations	el_list_conversations, el_get_conversation, el_delete_conversation, el_get_conversation_audio, el_get_signed_url, el_send_conversation_feedback
Knowledge base	el_list_kb_documents, el_get_kb_document, el_create_kb_document (url/text/file), el_delete_kb_document, el_get_kb_document_content, el_get_kb_dependent_agents, el_rag_index
Agent tools	el_list_tools, el_get_tool, el_create_tool, el_update_tool, el_delete_tool, el_get_tool_dependent_agents
Tests	el_list_tests, el_get_test, el_create_test, el_run_tests_on_agent, el_get_test_invocation
Telephony	el_list_phone_numbers, el_get_phone_number, el_import_phone_number, el_update_phone_number, el_outbound_call, el_submit_batch_call, el_list_batch_calls, el_get_batch_call, el_batch_call_action
Workspace	el_get_workspace_settings, el_update_workspace_settings, el_list_secrets, el_create_secret, el_delete_secret, el_list_llms, el_get_widget_config, el_search_voices
Escape hatch	el_api_request — any other /v1/convai/* endpoint (branches, versions, analytics, WhatsApp, MCP integrations, …)

Design notes:

• Safety: every destructive tool (delete_*, outbound/batch calls, workspace settings) requires confirm=true. DELETE is blocked on the raw escape hatch.
• Tool annotations: every tool ships MCP ToolAnnotations (readOnlyHint, destructiveHint, idempotentHint, openWorldHint) so clients can flag safe vs. side-effecting tools. Read-only = GET-only tools (26 of 50); the destructive ones line up with the confirm=true gates.
• Context-friendly: responses truncate at 50k chars; el_get_conversation supports transcript_only / include_transcript=false; voice search returns slim results.
• Data residency: point ELEVENLABS_API_BASE at your regional endpoint (EU: https://api.eu.residency.elevenlabs.io).

Setup

cd elevenlabs-agents-mcp
python3 -m venv .venv && .venv/bin/pip install -r requirements.txt

Claude Desktop / Cowork

Add to claude_desktop_config.json (Settings → Developer → Edit Config):

{
  "mcpServers": {
    "elevenlabs-agents": {
      "command": "/ABSOLUTE/PATH/elevenlabs-agents-mcp/.venv/bin/python",
      "args": ["/ABSOLUTE/PATH/elevenlabs-agents-mcp/server.py"],
      "env": { "ELEVENLABS_API_KEY": "sk_..." }
    }
  }
}

Claude Code

claude mcp add elevenlabs-agents \
  -e ELEVENLABS_API_KEY=sk_... \
  -- /ABSOLUTE/PATH/elevenlabs-agents-mcp/.venv/bin/python /ABSOLUTE/PATH/elevenlabs-agents-mcp/server.py

As part of a Cowork/Claude Code plugin

Reference it in the plugin's .mcp.json:

{
  "mcpServers": {
    "elevenlabs-agents": {
      "command": "python3",
      "args": ["${CLAUDE_PLUGIN_ROOT}/mcp/server.py"],
      "env": { "ELEVENLABS_API_KEY": "${ELEVENLABS_API_KEY}" }
    }
  }
}

Cloud Run deployment (production)

The repo ships production-ready: http_server.py (streamable HTTP transport, stateless JSON), Dockerfile, deploy.sh.

gcloud auth login && gcloud config set project YOUR_PROJECT_ID
ELEVENLABS_API_KEY=sk_... ./deploy.sh

The script enables APIs, stores the API key and a generated 64-char bearer token in Secret Manager, deploys from source to Cloud Run (Frankfurt europe-west3, 512Mi, scale-to-zero, max 3 instances), smoke-tests /healthz + auth, and prints the connect command:

claude mcp add --transport http elevenlabs-agents https://SERVICE_URL/mcp \
  --header "Authorization: Bearer TOKEN"

Production properties:

• Auth: every /mcp request needs the bearer token (constant-time compare); server refuses to boot without a ≥32-char token. Alternative IAM-only mode: MCP_ALLOW_UNAUTHENTICATED=true + --no-allow-unauthenticated.
• Stateless JSON transport: no sessions/SSE — safe for horizontal scaling and scale-to-zero.
• Secrets: only in Secret Manager, injected as env; never in image, logs, or responses.
• Logging: structured JSON per request (method/path/status/latency) → Cloud Logging. Bodies and transcripts are never logged (PII).
• Resilience: retry with exponential backoff + Retry-After on 429/502/503/504; pooled connections; 90s upstream timeout, 300s request timeout for long simulations.
• Read-only mode: deploy with READ_ONLY=true ./deploy.sh to hard-block all non-GET operations server-side (good for analytics-only access).
• Cost/blast-radius caps: min-instances=0 (≈€0 idle), max-instances=3.
• Token rotation: ROTATE_TOKEN=1 ./deploy.sh.
• Non-root container, slim base image, layer-cached deps.

Environment variables

Variable	Required	Default	Purpose
ELEVENLABS_API_KEY	yes	—	API key (Profile → API keys)
MCP_AUTH_TOKEN	HTTP mode	—	Bearer token protecting /mcp
MCP_ALLOW_UNAUTHENTICATED	no	false	Skip bearer auth (only behind IAM)
ELEVENLABS_MCP_READ_ONLY	no	false	Block all non-GET API calls
ELEVENLABS_API_BASE	no	https://api.elevenlabs.io	Data-residency region
ELEVENLABS_MCP_MAX_CHARS	no	50000	Response truncation limit
PORT	no	8080	HTTP port (Cloud Run sets this)

Quick test

ELEVENLABS_API_KEY=sk_... npx @modelcontextprotocol/inspector .venv/bin/python server.py

Example prompts once connected

• "List my ElevenLabs agents and show the config of the Projektinterview agent."
• "Duplicate agent X, change the system prompt to …, then simulate a conversation with a skeptical consultant persona."
• "Pull yesterday's conversations for agent X and summarize failure reasons."
• "Add https://example.com/faq to the knowledge base and enable RAG on it."