Text to Speech MCP Server — Hosted TTS for AI Agents (2026)

MCP (Model Context Protocol) is the standard AI assistants use to gain abilities: you give the assistant a server URL, and every tool that server exposes becomes something it can do. A TTS MCP server exposes speech tools — connect one and Claude, ChatGPT, Cursor, or any MCP-capable agent can turn the text it writes into audio. EasyVoice's is hosted: paste https://easyvoice.ae/api/mcp into your client with an API key, and audio comes back as hosted MP3/WAV download links — download promptly, links are not guaranteed to persist.

On claude.ai and Claude Desktop: Settings → Connectors → add a custom connector with the URL https://easyvoice.ae/api/mcp, then add the request header Authorization: Bearer ev_your_api_key (the API-key header option is labeled beta in the connector UI). In Claude Code it's one line: claude mcp add --transport http easyvoice https://easyvoice.ae/api/mcp --header "Authorization: Bearer ev_your_api_key". Keys are free at easyvoice.ae/account/keys — no card.

ChatGPT developer mode (web) accepts remote MCP servers, but its connector form has no custom-header field — so connect with the key-in-URL form https://easyvoice.ae/api/mcp?key=ev_your_api_key and authentication set to none. One hedge: community reports say ChatGPT Plus and Pro may restrict custom-connector tools to read-only, which would block text_to_speech there — this is multiply reported but not confirmed in OpenAI's docs, so treat write tools on Plus/Pro as unverified. Cursor connects via ~/.cursor/mcp.json with a url plus a headers entry carrying the Authorization header. Any stdio-only client can bridge with npx mcp-remote.

Claude can't synthesize speech by itself — you connect it to a text-to-speech service over MCP. Add EasyVoice's server as a custom connector (or via claude mcp add in Claude Code) with a free API key, then ask Claude to call the text_to_speech tool: it returns a hosted MP3 or WAV download link. A free key covers 12 voices and 5,000 characters/day, no card.

Yes, with a Pro key. The create_podcast tool takes alternating speaker A/B dialogue segments — up to 60 segments and 30,000 characters per episode — with a voice for each speaker. The job is queued, the agent polls get_job_status, and the completed job carries the stitched episode URL plus per-segment audio URLs. Free keys get a truthful Pro-required error instead.

Yes — enrollment, with guardrails, on a Pro or active 7-day pass key. The clone_voice tool requires a consent: true attestation (recorded server-side) and an audio sample passed as exactly one of audio_base64 or a public https audio_url — 10 MB max, wav/mp3/m4a, a 10–30 second clean recording (enrollment rejects shorter or longer clips). Account limits: 3 cloned voices maximum, 5 enrollments per day; cloned output carries an AudioSeal watermark, and cloned voices speak English. Over MCP, text_to_speech accepts your ready voice_* ids with Pro or active 7-day pass access (up to 2,000 characters synchronously; longer text queues with Pro, while pass holders must split it into calls of 2,000 characters or fewer). Use your ready clones in create_podcast (Pro) as speaker A or B. generate_long_form stays Kokoro-only.

Yes, with a Pro key. The generate_long_form tool accepts a script up to 500,000 characters (~9 hours of audio) in a single job — Kokoro narration voices only; Arabic ar_* and cloned voice_* voices are rejected by the long-form engine. The agent polls get_job_status: parts render progressively (N of M), and the download URL appears when the job completes.

A truthful error, nothing else. The job tools return the Jobs API's own message — "The Jobs API requires a Pro subscription — free API keys can use /v1/audio/speech." — as a tool error. clone_voice needs Pro or an active 7-day pass; without one it returns its own error. Nothing is queued and nothing is charged. Pro is $9.99/month flat; there are no promo codes to hunt for.

A free API key works today for sync text_to_speech: 12 voices, 20 requests/minute, 2 concurrent requests, and 5,000 characters/day shared with the web app — no card required. Pro is $9.99/month flat: all 66 voices, unlimited characters (fair use: 10M per rolling 30 days), and the cloning, podcast, and long-form tools over the same URL — with a 14-day money-back guarantee.

ElevenLabs now runs a hosted MCP server at https://api.elevenlabs.io/v1/mcp (you sign in with your ElevenLabs account); its earlier local stdio server was archived in August 2026. Its hosted docs list agent-management tools and text to speech, and ElevenLabs bills speech in metered credits on your plan (free 10K credits a month, then paid credit tiers). EasyVoice's MCP server is also hosted — a URL, not an install — but flat-rate at $9.99/month unlimited, and it adds long-form narration, two-speaker podcast episodes and consent-verified voice cloning over the same URL. The ElevenLabs details above were re-checked on 2026-09-28.

Yes — that's the core use case. Any MCP-capable agent or automation can call text_to_speech and get a hosted audio URL back: CI jobs that announce deploy status out loud, agent workflows that voice their summaries, apps that need spoken alerts. Because Pro is $9.99/month flat with unlimited characters (fair use: 10M per rolling 30 days), an agent calling the tool in a loop can't surprise-bill you.

Seven: text_to_speech (synchronous speech — up to 8,000 characters per call, 4,000 on Arabic ar_* voices), list_voices, get_usage, generate_long_form (Pro), create_podcast (Pro), clone_voice (Pro or 7-day pass), and get_job_status. Every tool returns a plain-text summary plus structured results, and audio always arrives as a hosted download URL rather than inline bytes.