Find a Voice, Generate, Transcribe: Fish Audio connects to agents via MCP

Fish Audio’s MCP server connects Claude, Codex, Cursor, and Windsurf to search for voices, generate audio, and transcribe files.

An MCP server connects Fish Audio’s voice features to Claude, Codex, Cursor, and Windsurf for voice search, audio generation, and file transcription.

Finding a voice, listening to samples, assigning it a script, or transcribing a recording can now be handled directly from an AI agent. Fish Audio has opened an MCP server that provides access to its voice library, speech synthesis, transcription, and credit tracking from the tools already used for writing and coding.

The connection runs through a single endpoint, `https://api.fish.audio/mcp`, compatible with Claude Code, Claude.ai, Cursor, Codex CLI, Windsurf, and other MCP clients. Authentication uses OAuth: a browser window opens to connect the user’s Fish Audio account, with no API key to copy into the configuration.

Once access is granted, requests can be made in natural language. The agent can find the most popular Japanese voices and play their samples, select a calm English voice to read a file, turn a script into audio, or transcribe a recording before extracting its action items. It can also check the account’s remaining credit balance.

In Claude Code or Codex, these features can be combined with terminal commands. An agent can read a folder of scripts, generate narrations in batches, download the resulting files, or upload a local recording to the transcription service. Generated audio is returned through permanent URLs. Local files are uploaded to temporary storage and automatically deleted after seven days.

Delivery can still be controlled from the text using instructions in square brackets, such as `[whispering]`, `[excited]`, or `[laughing]`. These tags shape the performance without being spoken aloud. If no voice is specified, the service selects a curated default suited to the requested language.

MCP server usage draws from the credits included in the user’s Fish Audio plan, just as it does in the web app. It does not consume credits from the developer API. The connector therefore offers another point of access: the API and SDKs remain designed for software integrations, while MCP brings voice features directly into conversations with the agent.