ChatGPT Skill — Find agentic resources (ARD)

ChatGPT Skill — Find agentic resources (ARD)

ChatGPT Skills are reusable, shareable instruction sets (beta; available on Business, Enterprise, Edu, Teachers, and Healthcare plans). Create a new Skill and paste the body below as its instructions.

Querying needs an HTTP capability. Pair this Skill with either an Agent Finder remote MCP connector (see mcp/chatgpt/) or a custom Action whose OpenAPI calls the Agent Finder POST /search endpoint. The Skill supplies the behavior; the connector or Action makes the call.

Skill name

agentfinder

Skill description (used for triggering)

Discover tools, skills, MCP servers, and agents for a task by searching ARD discovery services (Agent Finder). Use when the user wants to find a tool, skill, agent, MCP server, API, or capability. Offers a menu of named Agent Finders, remembers the choice, presents the ranked results, and never installs anything automatically.

Skill instructions

When the user asks you to find tools, skills, agents, MCP servers, or other capabilities for a task, follow these steps in order:

1. Pick an Agent Finder from a menu, and remember it. Agent Finders are listed — each with a name — in the shared agent-finders.json config (and, where the client can read/write files, in ~/.agentfinder/finders.json). Two ship by default: GitHub Agent Finder and Hugging Face Discover. The first time, show a numbered menu (name + description) and let the user pick; remember the choice — save it to ~/.agentfinder/finders.json's selected if you have file access, otherwise for the rest of the session — and don't ask again. The user can say switch agent finder to change. If you reach Agent Finder through a single fixed MCP connector or Action, that configured endpoint is the finder — skip the menu.

2. Query the chosen finder. Send an ARD search request to the selected finder's search URL: POST <search URL> with body { "query": { "text": "<the user's task>" } }. Add a filter (e.g. { "type": ["application/mcp-server+json"] }) to narrow.

3. Present the results as a numbered list: displayName, type, one-line description, publisher/identifier, endpoint URL, and relevance score. Say the score is relevance only — not a trust or safety rating. Offer to follow any referrals.

4. Never auto-install. Do not add, enable, connect, install, or invoke any returned resource yourself.

5. Install only on request. After the user picks one, give the steps to install or connect that resource themselves, then stop.