Cursor Text to Speech: Generate Voice-Overs From Your Editor

Cursor is an editor with an AI agent inside it. The agent already reads your code, edits files and runs commands. Add Clonylab as an MCP server and the same agent can also make voice-overs. You write a comment or a script, ask in the chat, and an MP3 lands in your project.

That is handy for more than you might think. Product demos need narration. Code walkthroughs are easier to follow when someone talks. Release notes can be read aloud for a team that prefers to listen. Once Cursor has the tools, all of that is one request away.

Clonylab gives the agent 14 tools for voices, speech jobs, History, Studio settings and usage. Our guide to the text to speech MCP server explains each tool. This page covers the Cursor part: the mcp.json file, the approval prompts and a few prompts that suit an editor.

Cursor reads its MCP servers from a file called mcp.json. There are two places for it. A file in your home folder, ~/.cursor/mcp.json, applies to every project. A file inside a project, .cursor/mcp.json, applies to that project only. For Clonylab we suggest the home folder, because your token should not sit in a project that other people can read.

Cursor’s documentation says the url and headers fields support config interpolation. In plain words, you can point at an environment variable instead of pasting the token. Set the variable once in your shell profile, then write this:

"headers": {
  "Authorization": "Bearer ${env:CLONYLAB_TOKEN}"
}

The token now lives in your shell settings, and the JSON can sit safely in a dotfiles repo. Start Cursor from a shell that has the variable, or set it where your system reads user variables, so Cursor can see it.

Open the agent chat. By default Cursor asks for approval before it runs an MCP tool, and you can open the arrow next to a tool name to see its arguments first. That is a good habit for the first few calls. Try these prompts, from small to useful.

Keep the voice the same across a project. Say it once in your first request, such as “use the calm US male voice for everything in this repository,” and Cursor will carry it through the chat. A consistent voice makes a set of clips feel like one product.

While the job runs, the agent can keep working on other steps. When the job is done, it saves the MP3 with a download command and tells you where the file is.

Cursor lets you decide how much the agent can do alone. Approvals are on by default, so a new tool call waits for you. Cursor’s Run Modes apply to MCP tools in the same way as to terminal commands. If you allow an MCP tool to run without asking, the agent can call it any time. For Clonylab, that means it can start voice-overs without a prompt.

Four Clonylab tools change something: generate_speech, cancel_job, delete_history and set_studio_settings. We suggest letting the read-only tools run freely and keeping approval on for delete_history, which removes voice-overs from your History for good.

The agent can pass four voice settings with each call. Each has a limit, and the Studio uses the same ones. Ask in plain words, such as “a little slower” or “a touch louder,” and the agent turns that into a number inside the range.

Want a single clip with full control? The Studio is made for that. For repeated work, such as a voice-over for every release, the editor route saves a lot of clicks. See also our guides for Claude Code and Windsurf.