Skip to content

Connect a chatbot

Use ARD directly from a chatbot to find tools, skills, MCP servers, and agents for a task — and decide what to install. The connectors repository provides ready-made setups for Claude, ChatGPT, GitHub Copilot, Microsoft Copilot, and Gemini.

The connectors are client-side only. A discovery service — such as GitHub's Agent Finder or Hugging Face Discover — exposes the search interface; the connector just points your chatbot at an endpoint you choose. Nothing is installed automatically.

What these connectors do

Once connected, ask your chatbot to find a capability for a task. These ready-made connectors will:

  1. Let you pick an Agent Finder to search — from a menu of named options (GitHub, Hugging Face, or your own), so you stay in control of where results come from. The Claude skill remembers your choice.
  2. Query it.
  3. Present a ranked list of matching resources — name, type, description, publisher, endpoint, and relevance score (relevance only, not a trust or safety rating).
  4. Never auto-install. Only after you pick a result does it show you how to add that resource yourself.

GitHub Copilot is the exception. Its skill defaults to GitHub's Agent Finder, so it skips step 1 and searches GitHub directly. The other four ask which finder to use.

This is one design among several

That flow is a deliberate choice — the most conservative one — not something ARD requires. ARD is just the discovery layer underneath; a connector can sit anywhere on a spectrum of autonomy:

  • (a) Find on request, then install on approval — what these connectors do. You ask it to find tools, it presents matches, and you choose what to install. A human is in the loop at both ends.
  • (b) Auto-discover, then install on approval — the chatbot decides on its own to query an Agent Finder when it hits a missing capability mid-task, but still asks before installing anything.
  • (c) Auto-discover and auto-install — the chatbot queries and installs without prompting. The most autonomous; appropriate only with a trusted, curated finder and a sandboxed environment.

These connectors ship as (a) for maximum user control, but you can build your own at any of these levels.

Two ways to connect

Method What it is Best when
Skill A portable instruction bundle — a Claude Skill, ChatGPT Skill, Gemini Gem, or Copilot agent — that drives the flow over the HTTP POST /search interface. Quickest start; works wherever the assistant can make an HTTP / Action call.
MCP Add Agent Finder as a remote MCP connector so the chatbot gets a native search tool. You want a first-class tool and your client supports MCP connectors.

Most setups use both: the MCP connector (or an Action) makes the actual call, and the Skill/instructions supply the "pick a finder, present, never auto-install" behavior.

How you add the Skill itself differs by client, and there is rarely a single "install" button:

  • Claude and GitHub Copilot — a SKILL.md file you drop into a skills folder (~/.claude/skills/, ~/.copilot/skills/), or install with a command.
  • ChatGPT — a Skill you add in the workspace UI (Business / Enterprise / Edu / Team plans).
  • Gemini — a Gem you create in the Gemini app.
  • Microsoft Copilot — an agent you build in Copilot Studio.

Each page below says exactly where it goes.

Pick your platform

Each page has the full install steps and how to invoke it, for both the Skill and the MCP connector:

  • Claude — Skill (Claude Code plugin) + remote MCP connector
  • ChatGPT — Skill (beta) + remote MCP (Developer mode)
  • GitHub Copilot — Agent Skill (SKILL.md) + remote MCP (VS Code mcp.json)
  • Microsoft Copilot — declarative agent + remote MCP (Copilot Studio)
  • Gemini — Gem + remote MCP (Gemini CLI)

Endpoints

The connectors ship with two named Agent Finders — GitHub Agent Finder and Hugging Face Discover — and you can add your own in agent-finders.json. They let you pick which to search (the Claude skill shows a menu and remembers your choice in ~/.agentfinder/finders.json). GitHub Copilot is the exception: it defaults to GitHub's Agent Finder. The examples below use the two shipped services interchangeably:

Discovery service Search endpoint
GitHub Agent Finder https://agentfinder.github.com/api/v1/search
Hugging Face Discover https://huggingface-hf-discover.hf.space/search

Point at either of these — or any compliant ARD discovery service — when a step asks for an endpoint. The connector never queries an endpoint you didn't choose.