Coding agents often generate calls against outdated endpoints or deprecated model IDs. This shows a three-layer fix, llms.txt for navigation, an MCP server for live lookup and skills for offline use, that API providers can copy and developers can point their agents at.
Key takeaways · AI-distilled
Its llms.txtA file at a site's root that gives AI systems a curated, plain-text map of its most useful content — robots.txt's counterpart for comprehension rather than permission.Full definition → is a curated Markdown index in six categories, opening with an agent quickstart and linking only to .md pages, while llms-full.txt carries fuller content, so an agent starts from a low-token map and escalates only when a task needs more.
A read-only Docs MCP serverA program that provides tools, resources and prompt templates to an MCP client; it may run on the same machine as the application or remotely.Full definition → at docs.fish.audio/Model Context Protocol (MCP)A published agreement for an AI application to discover what a separate program offers and ask it to run one of those offerings.Full definition → answers live questions such as current models, rate limits or emotion tags. In Claude Code it is added with claude mcp add --transport http, and no API key passes through the connection.
The installable fish-audio-api agent skillA reusable instruction file that teaches an agent how to do one job well — the procedure, the tools, and what counts as done.Full definition → (via npx skills, which supports 50+ agents) bundles auth, endpoint, encoding and WebSocket conventions for offline use, but it is fixed at install time and needs npx skills update to refresh.
Fish Audio's rule of thumb: use the skill for repeatable, offline code generation on known patterns, and the MCP server for open-ended questions and changes since install. It recommends both in production, with llms.txt as the first fetch.
Terms in this piece · Glossary
llms.txt — A file at a site's root that gives AI systems a curated, plain-text map of its most useful content — robots.txt's counterpart for comprehension rather than permission.
MCP server — A program that provides tools, resources and prompt templates to an MCP client; it may run on the same machine as the application or remotely.
Model Context Protocol (MCP) — A published agreement for an AI application to discover what a separate program offers and ask it to run one of those offerings.
agent skill — A reusable instruction file that teaches an agent how to do one job well — the procedure, the tools, and what counts as done.