agentfinder

Find agentic resources (ARD)

Invoke this skill as /agentfinder <query>, where <query> is the task the user wants to find resources for. Also use it whenever the user otherwise asks you to find tools, skills, agents, MCP servers, or other capabilities for a task. It searches ARD discovery services (Agent Finders) and presents matches for the user to choose from.

Requirements. Querying a finder needs an HTTP capability — in Claude Code, Bash with curl; or an Agent Finder MCP connector (see mcp/claude/ in this repo); or a fetch/web tool. If none is available, tell the user and point them at the MCP connector setup.

Follow this contract exactly:

1. Choose an Agent Finder (a menu the user sees only once)

Agent Finders are listed in a shared config at ~/.agentfinder/finders.json, each with a name. The user's choice is remembered there, so this is a one-time menu — not a question on every search.

1. Seed it if missing. If ~/.agentfinder/finders.json does not exist, create the directory and write this default:

   {
     "selected": null,
     "finders": [
       {
         "id": "github",
         "name": "GitHub Agent Finder",
         "description": "GitHub's public catalog of installable MCP servers, skills, and tools.",
         "search": "https://agentfinder.github.com/api/v1/search",
         "mcp": "https://agentfinder.github.com/api/v1/mcp"
       },
       {
         "id": "huggingface",
         "name": "Hugging Face Discover",
         "description": "Hugging Face's discovery service for agentic resources.",
         "search": "https://huggingface-hf-discover.hf.space/search",
         "mcp": "https://huggingface-hf-discover.hf.space/mcp"
       }
     ]
   }
   

2. Use the saved choice. Read the file. If selected names a finder, use it without prompting — say once: "Searching the saved finder's name — say switch agent finder to change." Then go to step 2.

3. Otherwise, show the menu. Present the finders as a numbered list (name + description) and let the user pick by number or name:

   Which Agent Finder should I search?
     1. GitHub Agent Finder — GitHub's public catalog of MCP servers, skills, and tools
     2. Hugging Face Discover — Hugging Face's discovery service
   (Add your own in ~/.agentfinder/finders.json.)
   

   Save the pick by writing its id to selected in the file, then continue.

When the user says switch agent finder (or similar), re-show the menu and update selected. If you have no file access (e.g. claude.ai or Desktop over the MCP connector), there's nothing to choose — just search the endpoint that connector is configured with.

2. Query the chosen Agent Finder

POST <the selected finder's "search" URL>
Content-Type: application/json

{ "query": { "text": "<the user's task, in plain language>" } }

Narrow results with a filter when useful — e.g. MCP servers only:

{ "query": { "text": "<task>", "filter": { "type": ["application/mcp-server+json"] } } }

3. Present the results

Numbered list. For each result: displayName, type, a one-line description, the publisher / identifier, the endpoint URL, and the relevance score. State that the score is relevance only — not a trust or safety rating. Offer to follow any referrals to other discovery services.

4. Never auto-install

Do not add, enable, connect, install, or invoke any returned resource yourself. Installation is always the user's explicit choice.

5. Install only on request

Once the user picks a result, give them the steps to install or connect that resource themselves (add it as an MCP connector, install the skill, or call its API) using the resource's own endpoint and protocol. Then stop and let them act.

Installation

Claude Code (recommended) — add this repo as a plugin marketplace and install:

/plugin marketplace add ards-project/connectors
/plugin install agentfinder@ard-connectors

Manual — copy this agentfinder/ folder into your Claude Code skills directory: ~/.claude/skills/ (personal) or .claude/skills/ (project).

Custom skills are currently Claude Code–only. The claude.ai web app and Claude Desktop do not yet support uploading your own skills.