Client-side connectors (Skills + MCP) that let an AI client discover agentic resources through ARD (Agentic Resource Discovery): the MCP servers, skills, agents and APIs that can actually do a task.
Upstream project: ards-project/ard-connectors. This distribution is maintained by Neuronto and differs in two ways: it ships as an installable Claude Code plugin marketplace, and the Agent Finder list contains every public ARD discovery service rather than a subset, so the menu reflects the whole ecosystem.
Named
neurontorather thanard-connectorson purpose. Claude Code registers one marketplace per name and a second one with the same name replaces the first, so reusing upstream's name would silently evict it for anyone who had both installed. The plugin isneuronto-agent-finderfor the same reason: the bareagentfindername is upstream's and is what their site promotes.
/plugin marketplace add neuronto/ard-connectors
/plugin install neuronto-agent-finder@neuronto
The plugin bundles the neuronto MCP connector, so the skill can query the
federated index straight away with no further setup. Removing or replacing it is
a one-line edit to .mcp.json.
Then ask for what you need in plain language:
/agentfinder a tool that can read PDFs
The skill shows the Agent Finder menu once, remembers your choice in
~/.agentfinder/finders.json, presents ranked results, and never installs
anything for you. Installation is always your explicit step.
MCP connector configs, one per client, in mcp/:
| Client | File |
|---|---|
| Claude Desktop | mcp/claude/claude_desktop_config.json |
| Cursor | mcp/cursor/mcp.json |
| GitHub Copilot / VS Code | mcp/copilot/mcp.json |
| Gemini | mcp/gemini/settings.json |
| Google Antigravity | mcp/google-antigravity/mcp_config.json |
| ChatGPT | mcp/chatgpt/README.md |
Client instruction files (rules, gems, custom instructions) are in
skills/ for clients that do not support MCP.
Each config points at a single endpoint. Change the URL to any finder in
agent-finders.json; the request shape is the same,
because they all implement the same ARD search interface.
Every public ARD discovery service known at the time of writing, alphabetical,
none preselected. Full details in agent-finders.json.
| Finder | Search endpoint | Notes |
|---|---|---|
| ARD Registry | ardregistry.org/api/search |
serves search at /api/search, not the spec's /search |
| Desvela ARD Registry | registry.desvela.dev/search |
also offers domain preflight tooling |
| GitHub Agent Finder | agentfinder.github.com/api/v1/search |
GitHub's catalog of installable resources |
| Hugging Face Discover | huggingface-hf-discover.hf.space/search |
indexes Hugging Face Spaces, Skills, MCP cards |
| Neuronto | neuronto.com/search |
federated: queries the others and fuses the rankings |
| WellKnown | wellknownhq.com/registry/search |
census of well-known manifests |
Add your own by appending to the finders array.
Every finder answers the same POST:
POST <finder search URL>
Content-Type: application/json
{ "query": { "text": "read a PDF and extract tables" } }Narrow by resource type:
{ "query": { "text": "...", "filter": { "type": ["application/mcp-server+json"] } } }Results carry a score. It is relevance only and is not a trust, safety or
quality rating. Do not present it as one.
Fixes that belong to everyone should go to upstream. Issues specific to this distribution, or a new Agent Finder to list, are welcome here.
Unresolved, and stated plainly rather than guessed. The upstream repository
publishes no licence: no LICENSE file, no licence field, no statement in its
README. Under default copyright that means no redistribution terms are granted
for the files inherited from it, so this fork cannot assert any either. (The ARD
specification repository is Apache-2.0; the connectors repository is not.)
Changes made in this fork are offered under Apache-2.0 by their author. That does not extend to the inherited files. Asking upstream to add a licence is the clean fix and is worth doing for everyone who forks it, not only for us.