AIエージェントガイド

You are helping a developer integrate Cartesia (Sonic TTS, Ink STT, voices, and related APIs).

Rules (required):
1. Fetch https://docs.cartesia.ai/llms.txt before guessing endpoints, parameters, or examples. Do not invent API fields or voice IDs.
2. Follow https://docs.cartesia.ai/ja-jp/use-the-api/api-conventions — send Cartesia-Version on every request.
3. Base URL: https://api.cartesia.ai (WebSockets: wss://). HTTPS only.
4. Default Cartesia-Version: 2026-03-01 unless the user specifies another date they tested with. On browser WebSockets, use ?cartesia_version=... (query wins over header when both are set).
5. Server/backend: Authorization: Bearer <api_key>. Create keys at https://play.cartesia.ai/keys
6. Browser/mobile: never embed API keys. Backend mints a short-lived access token (POST /access-token with a JSON body — {} or {"grants":{"tts":true}} — see https://docs.cartesia.ai/get-started/authenticate-your-client-applications). Client calls Cartesia with Authorization: Bearer <access_token> (or ?access_token= on WebSockets).
7. Web "button → hear audio": backend access-token endpoint → browser POST /tts/bytes with the token → response blob → Audio.play(). Examples: https://docs.cartesia.ai/ja-jp/examples/tts-play-audio and https://docs.cartesia.ai/ja-jp/examples/nextjs
8. Prefer official SDKs for app code: https://github.com/cartesia-ai/cartesia-python and https://github.com/cartesia-ai/cartesia-js
9. Sonic speed, volume, emotion: use generation_config (e.g. generation_config.speed), not deprecated top-level speed on REST payloads — https://docs.cartesia.ai/build-with-cartesia/capability-guides/volume-speed-emotion
10. Errors: for Cartesia-Version 2026-03-01 and newer, expect structured JSON — https://docs.cartesia.ai/use-the-api/api-errors
11. Cartesia Line (deployed voice agents, cartesia deploy, telephony) is a separate product — https://docs.cartesia.ai/line — not the same as calling TTS/STT REST APIs from the user's server.
12. CORS: browsers can call https://api.cartesia.ai directly with an access token in typical setups; proxy TTS through the user's backend if the client cannot reach Cartesia (corporate network, custom security policy, etc.).

Optional (often works better in the IDE):
- MCP (Cursor, Claude Desktop, etc.): https://docs.cartesia.ai/ja-jp/tools/ai/mcp — list voices, TTS/STT, dictionaries without custom scripts.
- Agent skills: run `npx skills add cartesia-ai/skills` and choose cartesia-api and/or line-voice-agent — https://docs.cartesia.ai/ja-jp/tools/ai/agent-skills