FindAgent documentation

Every agent on FindAgent is one package, published by a creator and ready to use in your LLM client. An agent is one package of up to four parts: Instructions, Skills, Actions and Code.

• Instructions — How the agent behaves: a prompt and example flows your AI app follows.
• Skills — Know-how your AI app loads when a task needs it, each skill served as a prompt.
• Actions — Calls to real tools with your own key, sent only to the hosts the agent declares.
• Code — A creator's own code from a GitHub repo, run in an isolated sandbox after review, or on your machine with your consent. The Code part →

Instructions, Skills and Actions run no code of their own; only the Code part does. Every agent page shows the parts it carries. Two things sit beside agents rather than inside them:

• MCP servers — An MCP server listing points at a server someone else runs; FindAgent lists it and never runs it. They live in the MCP directory; new to the term? See what an MCP server is.
• Departments — A Department is two to sixteen agents (up to eight for peer-to-peer) composed into one team that runs as a single MCP server. It is not a plugin. Departments →

You browse the catalogue on two axes — Industry (the vertical an agent serves) and Discipline (the kind of work it does, like Customer Service or Software Development). Lead with whichever matches how you think, then intersect the two from the Browse page.

An agent with actions or skills declares tools in its manifest. Each tool describes what it does and how it is invoked (an HTTP call, a prompt template, or a call into another agent). A shared, trusted runtime executes those declared tools, so the catalogue stays safe to use while agents still do real work. See the security model for why this matters.

• Connect. Add the agent's gateway address in any MCP client and sign in once. Everything runs on FindAgent, with nothing to download. A hosted agent's address looks like https://mcp.findagent.cloud/agents/<slug>; it works from web and mobile clients too. Connect guide →
• Install. Add it as a plugin from the catalog, in apps that install plugins. Parts that run on your own machine are Install-only. Install guide →

New to any of the words above? Each of these answers one question, in plain language:

• What is an MCP server? — A plain-language definition of the Model Context Protocol and MCP servers, how they differ from a plugin or a raw API, and how they connect to Claude, ChatGPT, Gemini, and Cursor.
• What is an AI agent? — What separates an agent from a chatbot or a prompt: tools, a goal, and the ability to take a multi-step action on your behalf.
• What is an agent marketplace? — How an agent marketplace differs from a plugin store or a prompt library, and what cross-LLM means for the agents you install.
• What can agents do? — The real tasks agents with actions handle today — analysis, support triage, reporting, repo work — and where the honest limits are.
• Cross-LLM, explained — Why an agent that works in Claude should also work in ChatGPT, Gemini and Cursor, and what has to be true for that to hold.

• Combine several agents into a team with an agent-to-agent topology and run them as one unit — see Departments.
• Publishing your own agent? Start with the creator guide and the manifest spec. You can even submit straight from your AI client — see MCP & API.
• Looking for a remote MCP server? Browse the MCP directory and connect any listing to your client.
• Integrating programmatically? See the API & integration reference.