What StealthStack is: an ARD registry

ARD went public on June 17, 2026, and within days we had started building StealthStack. The idea behind it was plain. If businesses post their agent tools in small files on their own websites, someone has to read all those files and make them searchable for anyone. As we did that, we kept running into jobs with no good tool behind them, so we started filling some of those gaps ourselves.

Why we built an ARD registry

ARD, or Agentic Resource Discovery, is an open specification that lets a business post a list of its agent tools on its own website. It's still young: the current version is 0.91, marked as a proposal that is still changing. If you are new to it, start with what ARD is and how it works. The short version is that the lists are spread across many websites, and an agent can't use a list it never sees.

The spec calls the service that collects those lists a registry. It works like a web search engine that only indexes agent tools, and it is what we set out to build. We call it an honest search of the agentic web, because we would rather show a weak result than dress it up.

We are an independent implementer. We follow the ARD spec, but we did not write it and we don't speak for the working group that did.

How our ARD registry works

Our crawler visits websites and reads the ARD file each one keeps at a standard address. It also follows links from one list to another, and from other public directories. Everything it collects goes into one index, and each result shows what the tool does, which website published it and how to connect. Skills, MCP servers, agents and web APIs all sit side by side, so one search covers them all.

We don't hand-pick who gets listed. If our crawler finds a site's ARD file, the listings go in, even ones that fail our name check, which we label. Listings that reach us through another directory show which directory relayed them, so you can trace each one back to its source.

People search the index on our search page. Agents can search it through our MCP server or through a web API. MCP is an open connection standard, first announced by Anthropic in November 2024, and assistants such as Claude support it. Our docs explain both.

The gaps we saw

We learned three things quickly.

First, most AI assistants have no built-in way to search ARD yet. The spec is new, and support will take time. So we built our MCP server as a bridge: add it once, and an assistant can search every list we have read.

Second, most of what is published today was built by developers for developers. Developers were the first to build agent tools, and they built them for their own work. Tools for plain business tasks exist, but they are still the smaller group, and some everyday jobs have very little behind them yet.

Third, some everyday business jobs had few good options in the index. The one we kept seeing was old files. Offices hold years of records in formats that newer tools and AI assistants often can't read: spreadsheets from the early 2000s, saved email, calendar exports, map files from surveyors. So we wrote tools for them. We now publish 30 skills (September 2026). Each one bundles written instructions with the code that does the work, in the format Anthropic describes for agent skills. All but one also run on our own servers, which spares an agent any local setup. We add a skill when a common file type has no good tool in the index.

Two kinds of result

Our results come in two kinds: tools that other makers publish, and a small set of our own skills that we also run on our servers for the gaps we saw.

Most results come from other makers. When your agent picks one of those, it connects straight to that maker's tool. We are not in the middle, we don't see the work and we don't sell access.

Some results are our own skills. When your agent picks one of those, it can load the skill and run it locally. It can also call our hosted version, and then the work runs on our servers.

A note on trust

For listings we find on a publisher's own site, we compare the web address in the listing's name with that site. Most of those match, which confirms they came from the publisher's own website. Others were copied from other directories, so we pass them on without that check. A few don't match at all and are labeled "domain mismatch" on the site, so treat those with extra care. The check tells you where a listing came from, not whether the tool is safe. As the ARD FAQ puts it, ARD's job is "to communicate trust, not to confer it." Our docs have the details.

Next steps

1. Connect your assistant to our index using the steps in our docs.
2. Search the index for a task from your week, and look at which website published each result.
3. See what we publish ourselves in our skills today, and what comes next. If you are weighing whether this matters for your firm, AI agents for small business makes that case.

Frequently asked questions

Is StealthStack free to use?

Yes. Search, the MCP server and the web API are free today and need no account or key.

Does StealthStack own or run ARD?

No. ARD is an open specification from a separate working group, and StealthStack is not a member. We build a registry that follows the spec.

Why don't I see my business in the index?

We only list what a business publishes in an ARD file on its own site or through another directory we read. If you have no ARD file yet, there is nothing for us to find. The ARD explainer covers what goes into one and who usually sets it up for a small business.