Agentic Resource Discovery · now an IETF Internet-Draft
Neuronto is a federated ARD registry. Ask it what can do a task and it searches its own catalogue and every other public ARD registry at once, then fuses the rankings into a single answer. It currently indexes 177,024 resources from 98,447 publishers, 162,313 verified to respond.
ARD stands for Agentic Resource Discovery. It is an open specification for how AI agents find the tools, skills, agents and APIs they need, published in June 2026 by a working group including Google, Microsoft, Hugging Face, AWS, Cisco, GitHub, Nvidia, Salesforce and Snowflake.
An agentic resource is anything an AI client can call to get work done: an MCP server, an A2A agent, a skill, an API, a workflow. ARD answers one question, “what is available for this task?”, and then gets out of the way. It is a discovery layer, not a runtime, and it does not replace MCP or A2A: it tells an agent what exists, and the agent connects using the resource's own protocol.
An agent can only use capabilities someone wired in beforehand, and every tool description competes for space in the context window. That does not survive an ecosystem of thousands of tools. ARD moves selection out of the context window and into a search service.
Publishers describe what they offer at /.well-known/ard.json on their own domain. Registries crawl those manifests and answer queries over them. Nobody has to be approved by a marketplace to be found.
The specification defines three federation modes and makes auto the default: a registry queries its peers, merges their results and returns one set. Ask once, reach the whole federation.
Agent builders who want to stop hard-coding integrations. API and tool vendors who want to be discoverable without applying anywhere. Platform teams running discovery over their own internal services.
Every entry is a resource an AI client can actually call, held with its endpoint, its description, the queries its publisher says it serves, and which sources carry it. The whole index is queryable over the API with no key and no signup, and browsable page by page through GET /agents.
| Kind of resource | Entries | What it is |
|---|---|---|
| mcp-server | 159,688 | MCP servers an agent can connect to and call as tools |
| skill | 11,682 | Skills and agent skill bundles |
| openapi | 2,996 | REST APIs described by an OpenAPI document |
| other | 886 | Resources whose type is outside the named set |
| doc | 753 | Machine-readable documentation such as llms.txt |
| plugin | 565 | Plugin manifests, including the OpenAI-era ai-plugin.json |
| a2a-agent | 326 | A2A agents with a published agent card |
| catalog | 97 | Nested catalogues pointing at further entries |
| registry | 18 | Other ARD registries, which is how federation is discovered |
| package | 7 | Published packages |
| graphql | 2 | GraphQL APIs |
| webmcp | 2 | Tools a web page exposes to a browser agent, W3C WebMCP |
| agent | 1 | Agent descriptors that are not A2A cards: ACP, OASF, AgentFacts |
| dataset | 1 | Published datasets |
Aggregated from every public ARD registry, the official MCP Registry, the major community aggregators, and our own crawl of the well-known discovery paths. No single source is close to complete on its own: the official registry does not carry servers their authors never submitted to it, and each aggregator holds thousands the others do not, so the table below is the union rather than any one of them. The crawl reaches publishers no other index covers: the only other general crawler in the ecosystem stops at the top 100,000 domains, so anyone smaller is invisible everywhere else.
| Source | Entries |
|---|---|
| glama | 68,113 |
| Official MCP Registry | 58,255 |
| mcpso | 18,482 |
| smithery | 10,239 |
| GitHub Agent Finder | 8,874 |
| WellKnown | 5,168 |
| Our own crawl of the discovery paths | 5,020 |
| Hugging Face Discover | 3,087 |
| apis.guru | 2,509 |
| ardregistry | 1,684 |
| submitted | 919 |
| Desvela | 720 |
| dockerhub | 100 |
| Publisher | Entries | Verified live | Kinds |
|---|---|---|---|
mcp.so mcp-server | 18,482 | 18,482 | 1 |
github.com mcp-server · openapi · plugin | 8,104 | 8,104 | 3 |
huggingface.co mcp-server · registry · skill | 3,065 | 2,962 | 3 |
io.github.sadri-dridi mcp-server | 2,366 | 2,366 | 1 |
io.github.pipeworx-io mcp-server | 1,916 | 1,856 | 1 |
pipeworx.io mcp-server | 1,900 | 1,885 | 1 |
ora.ai a2a-agent · mcp-server · skill | 1,666 | 1,653 | 3 |
pipeworx-io mcp-server | 1,390 | 1,390 | 1 |
io.github.mcp-dir mcp-server | 1,115 | 1,115 | 1 |
mcp-dir mcp-server | 1,113 | 1,113 | 1 |
io.github.mrfentmen mcp-server | 1,079 | 214 | 1 |
registry.modelcontextprotocol.io mcp-server | 777 | 777 | 1 |
Benchmarked 31 August 2026 with the specification's own conformance tool. Precision is top-1 over six queries with unambiguous answers; spread is the average gap between the best and worst score in a result set.
| Registry | Conformance | Entries | P@1 | Spread | Latency | federation: auto |
|---|---|---|---|---|---|---|
| Neuronto | PASS | 177,024 | 6/6 | 78 | <200ms | yes |
| GitHub Agent Finder | FAIL | 344 | 6/6 | 32.5 | 1,970ms | no |
| WellKnown | PASS | 1,239 | 6/6 | 6.7 | 401ms | no |
| Hugging Face Discover | PASS | 412 | 0/6 | 40.8 | 1,712ms | no |
| Desvela | PASS | 93 | 2/6 | 39.7 | 213ms | no |
curl -s https://neuronto.com/search \
-H 'content-type: application/json' \
-d '{"query":{"text":"scrape a website"},
"federation":"auto"}'
claude mcp add --transport http \
neuronto https://neuronto.com/mcp
Publishing? Serve an ARD manifest at /.well-known/ard.json with representativeQueries on every entry. The crawler finds it. No submission form, no allowlist.