generated: '2026-09-17'
method: derived
source: |
  MCP side — the provider's documented capability list at
  https://elevenlabs.io/docs/eleven-agents/operate/hosted-mcp (the live tools/list is
  OAuth-gated: POST https://api.us.elevenlabs.io/v1/mcp returned HTTP 401 with an RFC 9728
  challenge on 2026-09-17).
  REST side — openapi/elevenlabs-openapi.json (the provider's published OpenAPI 3.1.0,
  301 paths / 390 operations, fetched from https://api.elevenlabs.io/openapi.json).
description: |
  Binding between the hosted MCP server's documented capabilities and the REST operations
  that back them.

  CONFIDENCE CAVEAT, read this first: the hosted server's real tool names and inputSchemas
  could not be read — anonymous tools/list is OAuth-gated. The left-hand column is therefore
  the PROVIDER'S PROSE description of each capability, given a descriptive slug by us, NOT a
  tool identifier returned by the server. Bindings are by semantics against operationIds
  verified to exist in the published spec. No confidence above `medium` is claimed for any
  row, because the mapping cannot be confirmed until an authenticated introspection runs.
  Re-run this after an authenticated tools/list and every row can be settled.
surfaces:
  openapi:
    path: openapi/elevenlabs-openapi.json
    url: https://api.elevenlabs.io/openapi.json
    operations: 390
    gated: false
  mcp:
    url: https://api.elevenlabs.io/v1/mcp
    regional:
    - https://api.us.elevenlabs.io/v1/mcp
    - https://api.eu.residency.elevenlabs.io/v1/mcp
    - https://api.in.residency.elevenlabs.io/v1/mcp
    - https://api.sg.residency.elevenlabs.io/v1/mcp
    gated: true
    gate: OAuth 2.0 bearer (RFC 9728 protected resource)
  graphql:
    url: null
    gated: null
    note: ElevenLabs publishes no GraphQL endpoint. The repo's graphql/ directory holds an
      API Evangelist analysis document, not a provider schema.
crosswalk:
- tool: create_agent
  capability: Create new agents by describing what you want
  category: agents
  rest:
  - create_agent_route
  binding: direct
  confidence: medium
  note: POST /v1/convai/agents/create. The MCP tool takes a natural-language description, so
    the server almost certainly composes the AgentConfig body rather than passing it through.
- tool: update_agent
  capability: Update any agent setting, including the system prompt, voice, language, and
    first message
  category: agents
  rest:
  - patch_agent_settings_route
  binding: direct
  confidence: medium
- tool: list_agents
  capability: List your agents and inspect or compare their configurations
  category: agents
  rest:
  - get_agents_route
  - get_agent_summaries_route
  - get_agent_route
  binding: composite
  confidence: medium
  note: list + per-agent get; "compare" implies the client fans out over get_agent_route.
- tool: duplicate_agent
  capability: Duplicate agents
  category: agents
  rest:
  - duplicate_agent_route
  binding: direct
  confidence: high
  note: POST /v1/convai/agents/{agent_id}/duplicate is a 1:1 match for the documented verb.
- tool: delete_agent
  capability: Delete agents
  category: agents
  rest:
  - delete_agent_route
  binding: direct
  confidence: high
  note: DESTRUCTIVE and irreversible — no restore operation exists. The provider itself
    flags this tool for human approval. See conventions/elevenlabs-conventions.yml
    reversibility.irreversible.
- tool: list_conversations
  capability: Review an agent's recent conversations
  category: conversations
  rest:
  - get_conversation_histories_route
  binding: direct
  confidence: medium
- tool: get_conversation_transcript
  capability: Read full transcripts
  category: conversations
  rest:
  - get_conversation_history_route
  - get_conversation_summary_route
  binding: composite
  confidence: medium
- tool: get_agent_topics
  capability: Explore the topics your agents' conversations cover
  category: conversations
  rest:
  - get_agent_topics_route
  binding: direct
  confidence: high
  note: GET /v1/convai/agents/{agent_id}/topics — the only operation in the spec that
    produces conversation topics.
- tool: estimate_llm_usage
  capability: Estimate an agent's expected LLM usage and cost before making changes
  category: analytics
  rest: []
  binding: unmapped
  confidence: low
  note: no operationId in the published spec obviously produces an LLM cost estimate. Listed
    here rather than in mcp_only because the backing operation may exist under a name this
    pass did not resolve. Settle with an authenticated tools/list.
- tool: get_agent_widget
  capability: Retrieve an agent's widget configuration and shareable link
  category: agents
  rest:
  - get_agent_widget_route
  - get_agent_link_route
  binding: composite
  confidence: high
  note: two operations, exactly matching the two things the capability names.
- tool: get_knowledge_base_size
  capability: Check the size of an agent's knowledge base
  category: knowledge-base
  rest: []
  binding: unmapped
  confidence: low
  note: the knowledge-base surface has 31 operations; none is named for a size/quota read.
    Likely computed server-side from the document list.
- tool: text_to_speech
  capability: Generate speech audio from text, returned as a short-lived download link
  category: generation
  rest:
  - text_to_speech_full
  binding: adapted
  confidence: medium
  note: the REST operation returns an audio stream; the MCP tool returns a short-lived
    DOWNLOAD LINK instead, because an MCP transport cannot hand an agent a binary body. A
    real shape difference, not a naming difference.
mcp_only:
- tool: estimate_llm_usage
  reason: no matching public REST operation identified; may be an MCP-side composition.
- tool: get_knowledge_base_size
  reason: no matching public REST operation identified; likely derived server-side.
rest_only_summary:
  total_rest_operations: 390
  operations_reachable_via_documented_mcp_capabilities: 15
  rest_only_estimate: 375
  note: |
    The hosted MCP server is deliberately narrow. It covers ElevenAgents management plus one
    generation verb; the other ~375 operations — Studio, dubbing, productions, voice cloning,
    music, speech-to-text, the whole workspace/administration surface — have no hosted MCP
    tool. The separate open-source LOCAL server (pypi elevenlabs-mcp) covers more of the
    generation surface; its tool list was not introspected in this pass.
rest_only_families:
- /v1/studio/* (Studio projects, chapters, snapshots)
- /v1/dubbing/* and /v1/productions/*
- /v1/voices/* (including PVC cloning)
- /v1/speech-to-text/*, /v1/speech-to-speech/*, /v1/audio-isolation/*
- /v1/music/*, /v1/sound-generation/*, /v1/text-to-dialogue/*
- /v1/workspace/*, /v1/service-accounts/*, /v1/usage/*
- most of /v1/convai/* beyond agents and conversations (tools, mcp-servers, phone-numbers,
  batch-calling, secrets, testing, triage-tickets)
coverage:
  mcp_capabilities_documented: 10
  crosswalk_rows: 12
  bound_to_rest: 10
  unmapped: 2
  high_confidence: 4
  medium_confidence: 6
  low_confidence: 2