Give Claude, Cursor and ChatGPT a voice with MCP

What the assistant can do once connected

• List and filter the voice catalogue - language, accent, pace, tier, licence.
• Render a line or a script to a file, in the format you ask for.
• Read back what it just wrote, which is the fastest proofread there is - a bad sentence is audible before it is visible.
• Draft, render and re-render inside one conversation.

Setting it up

Create an API key in Settings, add Naadly's MCP endpoint to your client's server list with the key, and restart the client. The tools appear with their own descriptions; nothing else is needed. The key carries your plan's limits, so an over-enthusiastic agent runs out of quota rather than running up a bill.

What to expect from an agent holding a voice tool

Given the tool, a model will use it enthusiastically and not always wisely, so it is worth knowing the failure modes before you wire it into anything that matters:

• It will pick a voice by name if you let it. Name the voice you want in the instruction, or tell it to filter by language and accent first and show you the shortlist.
• It will render the whole document when you asked about a paragraph. Say how much to read.
• It cannot hear the result. If a name is mispronounced it will not notice - put the correction in the workspace pronunciation list, where every render inherits it.
• Renders take seconds, so a long script inside a conversation will look like the assistant has stalled; ask for a queued job instead when the script is long.
• Every file it produces is watermarked and counted against your plan exactly as if you had rendered it yourself.

Sensible uses

Reading a draft aloud to catch clumsy sentences; generating placeholder voice-over while prototyping a UI; producing the audio for a demo script inside the same session that wrote it; letting a coding agent narrate a walkthrough of the change it just made.

Questions

Which clients does the MCP server work with?

Any MCP client - Claude Desktop and Cursor are the common ones. The protocol is the interface, not the client.

Do I need a paid plan for MCP?

Yes - it authenticates with an API key, which is a paid feature from $15 a month.

Can an agent spend my whole quota?

It can spend up to your plan's limit and no further; usage is visible per key on the usage page.