ElevenLabs Agents API + MCP vs Inworld Realtime API

ElevenLabs Agents API + MCP 92

Public OpenAPI at api.elevenlabs.io/openapi.json (25). llms.txt with Markdown copies of each page (10). The docs explain each agent setting with guidance on when to use it, such as turn timeout and soft-timeout fillers (16 of 20). The OpenAPI types every field, but conversation_config is a large nested object (12 of 15). The errors page lists codes, types and fixes, and every error carries a request_id (14 of 15). Weekly dated changelog entries and a versioned API (15).

Inworld Realtime API 60

No AsyncAPI or other machine-readable file for the socket, and the spec paths redirect (0). llms.txt, Markdown twins and a docs MCP server (10). The overview, OpenAI migration guide, providerData reference and the billing page's account of how a session is counted say what to use when (16 of 20). The reference lists every client and server event with field types and required marks, as text rather than a schema (12 of 15). JavaScript and Python example projects and an error event with type, code and param, without a list of codes (9 of 15). Dated release notes with affected clients and required actions, but no API version beyond /v1 (13 of 15). Total 60.

ElevenLabs Agents API + MCP 83

The hosted MCP server signs in with OAuth and asks for scoped access to agents and speech. API keys can be limited to endpoint groups, given a credit limit, owned by a service account, rotated, and are disabled if found on GitHub (30). Scoped keys and signed URLs for private agents keep the main key off clients (17 of 20). Agents pass caller speech to an LLM. The guardrails docs, generally available since 19 October 2026, describe a Manipulation guardrail that detects prompt injection in user turns and can end the conversation, included at no extra cost, and the MCP security page recommends Always Ask approval for tool calls. It is a setting the builder turns on, not a default we could confirm (12 of 15, was 5). Audit logs cover over 100 endpoints through GET /v1/workspace/audit-logs, but only on Enterprise, and every conversation keeps a transcript (12 of 15). security.txt valid to 6 October 2027 (re-read on 1 November) and HIPAA BAAs are available with zero retention mode. We didn't re-check certifications or a bug bounty (12 of 20). 30 + 17 + 12 + 12 + 12 = 83.

Inworld Realtime API 58

Workspace API keys sent as Basic credentials, deleted to revoke, with separate keys per environment advised. The CLI mints read-only keys unless --write is given, but write toggles cover only the Voice and Router APIs, so no key is limited to one product. Browsers use short-lived tokens minted server side. Session IDs, not secrets, go in the URL (25 of 30). Tools run in the client, so the developer approves side effects, with no approval step in the API (10 of 20). The agent listens to callers and passes their speech to an LLM, and no prompt-injection guidance was found (5 of 15). Usage per key and a Performance page with errors and throttling (8 of 15). SOC 2 Type II and HIPAA claimed and a vulnerability report form, with no security.txt or bug bounty, per the 4 November reading (10 of 20). Total 58. Zero data retention covers session content from the Developer plan up, but not requests routed to third-party LLM providers.

ElevenLabs Agents API + MCP 78

Closed service with clear terms and MIT SDKs (18 of 30). Conversation data is kept 2 years by default, set per agent in days, with per-agent zero retention mode and audio saving that can be turned off, and the statements agree (22 of 30). Dated changelog entries, but we didn't find a deprecation policy for the agents product (10 of 20). EU, India and Singapore residency endpoints are documented. We didn't check the subprocessor list this run (15 of 20).

Inworld Realtime API 63

Closed service with clear terms naming Theai, Inc. dba Inworld AI (15). The ZDR page says what a session keeps and that third-party LLM requests aren't covered, but the terms and privacy notice disagree on training (12 of 30). Release notes flag potentially breaking changes with the clients affected and what to do, without stating a notice period (14 of 20). US headquarters with EU and India residency on Enterprise, and no sub-processor list in the docs, while sessions may route to third-party LLM providers (10 of 20). Total 51.

ElevenLabs Agents API + MCP 85

Changelog entry on 26 October 2026, and @elevenlabs/client 1.32.1 and @elevenlabs/react 1.20.2 tagged on 29 October (30). Weekly entries, with 108 tags in the agents packages repository since 3 August (20). Weekly changelog and incident write-ups on the status page, support we didn't test (12 of 15 for a closed service). Python and JavaScript SDKs with over a million weekly downloads each per the 1 October check, plus mobile SDKs (15). MIT SDKs on GitHub (8 of 10). 30 + 20 + 12 + 15 + 8 = 85.

Inworld Realtime API 71

Release note of 16 October 2026 (cache_transcript), 21 days ago (30). Three dated Realtime entries in the 90 days from 8 August, on 14 September, 23 September and 16 October (20). Discord and support@inworld.ai, with no evidence of answers read (8 of 15). Works with OpenAI's Realtime clients and the CLI is current at 1.5.0, but there is no Inworld Realtime SDK (8 of 15). CLI on npm under MIT, CI not read (5 of 10). Total 71.

ElevenLabs Agents API + MCP 75

Read as a hosted service. Status page at status.elevenlabs.io (incident.io) with an ElevenAgents component group (Conversations, Telephony, Integrations, Other Agents APIs, Quality, RAG, UI) and dated component history (20). Re-read on 1 November from the history page, which dates each component's impact, as the ElevenLabs text-to-speech and voice cloning listings were read; the incidents feed was re-posted on 21 October and leaves older incidents out. In the 90 days from 3 August, ten incidents touched Agents components, all partial or under an hour. The longest, from 08:50 UTC on 13 October, a database behind the Agents platform in the US stalled, and the write-up says roughly 45 per cent of US calls failed and 7 per cent of API requests returned errors until 14:18 UTC, on seven Agents components marked partial outage. That is a share of requests in one region, read as partial, as hubspot-mcp's "many customers in North America" was and as the ElevenLabs text-to-speech and voice cloning listings read this incident, not every user in a region as in google-drive-api's regional major. The others are Inbound Twilio calls on 18 August (under 1 per cent), EU batch calling delays on 27 August, EU conversation history on 10 September, some calls failing to start on 18 and 19 September (1 hour 25 minutes), the EU platform unresponsive from 13:40 to 14:18 UTC on 28 September (38 minutes, one region), a usage datastore fault from 15:56 to 19:36 UTC on 29 September that caused latency and some failed requests on Agents conversations, STT and Creative products (partial outage), EU SIP calls with static IPs from 9:53 to 12:06 UTC on 6 October (one region, one call path) and SIP conversations from 14:21 to 15:15 UTC on 8 October. Minor only (20 of 30, was 5). Concurrency per plan is published, from 4 calls on Free to 40 on Business with burst to three times the cap (15). The errors page returns 429 with rate_limit_exceeded, concurrent_limit_exceeded or system_busy and tells callers to back off exponentially. No Retry-After header and no idempotency keys found (10 of 15). No SLA found for a self-serve tier, and Enterprise lists custom terms around SLAs (0). Agents is generally available (10). 20 + 20 + 15 + 10 + 0 + 10 = 75.

Inworld Realtime API 62

Hosted reading of the checklist. status.inworld.ai is a public status page (20), with history we couldn't read, as on inworld-tts (5). Concurrent sessions are published per plan, 10 on On-Demand up to 1,000 on Growth (15). The rate limits page gives exponential backoff with jitter and Retry-After on longer windows, and sessions emit rate_limits.updated, with nothing on reconnecting a dropped session (12 of 15). An SLA appears only in the Enterprise column, with no terms (0). No preview or beta label on the Realtime API (10). 20 + 5 + 15 + 12 + 0 + 10 = 62.

ElevenLabs Agents API + MCP 77

We couldn't count the hosted MCP server's tools, since the docs list about ten capabilities without a tool list, so we scored context cost on the API, where list endpoints take page sizes and cursors (18 of 25). Conversation lists page and filter (18 of 20). Typed error codes with suggested fixes (18 of 20). MCP clients can set each tool to ask first and the docs warn that deleting an agent is destructive, but we found no idempotency keys (8 of 20). Few required fields to create an agent and official SDKs for Python, JavaScript, React, Swift, Kotlin and Flutter (15).

Inworld Realtime API 68

Scored as an API. Conversation history defaults to the last 100 items, can be trimmed, and can be cached on Anthropic models, and every response.done carries a usage block (18 of 25). No lists to page. Output controls are modalities, turn detection settings, optional transcription and response.speak (12 of 20). Error events follow OpenAI's type and code shape, unsupported fields come back as non-fatal unsupported_field warnings, and item_not_found names the bad parameter (14 of 20). Tools run on the client and a follow-up response starts automatically after a tool result. Nothing covers side effects of a tool call cut off by an interruption (12 of 20). Works with OpenAI Realtime clients after changing the URL and key, with no Inworld Realtime SDK (12 of 15). Total 68.