ARD: the standard for agents to find your tools
Google's Agentic Resource Discovery spec lets AI agents discover and verify tools via an ai-catalog.json you host and own. Why it matters for builders.
MCP taught agents how to use a tool once you hand them the connection. Nobody solved how an agent finds the right tool in the first place — across companies, across the open web, without a hardcoded list. On June 17, Google announced Agentic Resource Discovery (ARD), an open spec aimed squarely at that gap.
What actually happened
ARD is Apache 2.0, developed under a Linux Foundation working group. It's backed by an unusually wide bench — Help Net Security reports collaborators including Microsoft, Cisco, Databricks, GitHub, GoDaddy, Hugging Face, Nvidia, Salesforce, ServiceNow, and Snowflake. When Google and Microsoft ship the same spec, it's worth reading.
Two primitives do the work. First, a catalog: an ai-catalog.json file you publish at a well-known path on your own domain, listing the capabilities you expose — MCP servers, A2A agents, OpenAPI tools, Skills. Because it lives on your domain, domain ownership becomes the identity anchor. Second, registries: search engines that crawl those catalogs and return ranked, verified matches to a natural-language query like "find a tool that checks order status."
The key design choice: ARD sits before invocation. It answers where a capability lives, which one to use, and whether it's safe to connect — then hands off to the resource's native protocol (MCP, A2A, whatever). It's not a replacement for MCP. It's the discovery layer MCP never had.
Why it matters for your business
Think about what "agents can find your tools" means for a small operator. Today, if you want an AI agent — yours or a customer's — to reach your booking system, your inventory API, or your quote generator, someone wires it up by hand. ARD is a bet that discovery becomes ambient: you publish a catalog you control, and agents find your capabilities the way search engines find your site. It's SEO for the agentic web, and the file lives on infrastructure you own.
We like that last part. The whole point of a standard hosted at your domain, under your cryptographic identity, is that no platform sits between you and the agents that call you. That's the same reason we build systems clients own outright. ARD is early — a spec, not a finished ecosystem — but it's the right shape: portable, open, and anchored to the one asset a vendor can't repossess, your domain.
Key takeaways
- Google announced ARD on June 17 — an Apache 2.0 spec for how AI agents discover and verify tools across the web
- Two primitives: an
ai-catalog.jsonyou host on your own domain, and registries that crawl and rank those catalogs - ARD handles discovery before invocation, then hands off to native protocols like MCP or A2A — it complements MCP, doesn't replace it
- It's backed by both Google and Microsoft (plus GitHub, Nvidia, Salesforce, and more), and it anchors identity to a domain you own
Want agents — yours or your customers' — to reach your systems? We build vendor-agnostic integrations on open standards you own, not on some platform's walled garden. See our approach or tell us what you'd expose.
Sources: Google Developers Blog, Help Net Security.
- #ai-agents
- #ard
- #mcp
- #interoperability
- #standards
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
GLM-5.2 open weights: the cost case for portability
Z.ai's open-weight GLM-5.2 matches frontier coding models at a fraction of the API price. The real story isn't the benchmark — it's that switching costs just dropped.
Read itAgentic loops: how to make an AI agent finish the job
Anthropic's new Claude Code loops guide names four loop types. The one that matters for operators: proactive loops that do real work while you sleep.
Read it