The tool that changes how an agent works here is recommend_voices: you describe a style and get ranked voice IDs, instead of needing one in hand before you can start. Cloning is deliberately short-lived — the sample must be WAV or MP3 under 25 MB, and you delete the voice when the job is done.
An MCP server for the Typecast text-to-speech API. It exposes the voice catalogue, the synthesis endpoints and quick voice cloning as tools, so an assistant can choose a voice and produce audio without you writing API calls.
- get_voices and get_voice — browse the V2 catalogue filtered by model or gender, or read one voice by ID
- recommend_voices — describe the style, mood or language you want in plain text and get candidate voice IDs back with scores
- text_to_speech — synthesise a line with a chosen voice and model
- text_to_speech_stream — synthesise through the streaming endpoint and save the concatenated audio
- text_to_speech_with_timestamps — synthesise and return timestamp alignment for caption generation
- clone_voice and delete_cloned_voice — build a temporary voice from a WAV or MP3 sample, then remove it; cloned IDs start with uc_
- play_audio — play a generated file on the machine running the server
- get_my_subscription — read the plan tier, credits used and total, and the concurrency limit
- search_documentation — search the Typecast documentation
A Typecast API key. Locally, clients launch it with uvx from the repository and read TYPECAST_API_KEY plus TYPECAST_OUTPUT_DIR (default ~/Downloads/typecast_output); it needs Python 3.10 or newer, and XDG_RUNTIME_DIR on Linux. A hosted Streamable HTTP endpoint takes the key in an X-API-KEY header — without one it exposes only search_documentation, and play_audio stays local-only.
One command plus a key — uvx --from git+https://github.com/neosapience/typecast-api-mcp-server.git typecast-api-mcp-server, then supply credentials