Agents Can't Find Their Own Tools — A Runtime Tool-Discovery Layer

The Problem

The way an AI agent gets its tools today is install-first, use-later. A developer hardcodes an MCP server URL into a config file, and only then can the agent use that tool. This breaks at scale, because there is no way for an agent to discover an unfamiliar tool dynamically while it’s working. The common workaround — dumping every tool description into the LLM’s context window — has hard limits. Token budgets cap it, and once you’re past a few dozen tools the model gets confused about which one to pick. So the agent’s capability is trapped inside the set of tools a developer imagined and wired up ahead of time. Ask it to “put this PDF into the accounting system” and if that accounting connector wasn’t installed in advance, the agent doesn’t know it exists. A human would just search and find it; the agent can’t. Tools themselves are exploding — thousands of MCP servers, AI Skills, and ML applications — yet the layer that lets an agent find and select the right one at runtime is missing.

Why Now

What creates the timing is a standard arriving. Hugging Face, together with Microsoft, Google, GoDaddy, and others, published a draft Agentic Resource Discovery (ARD) spec. The model is simple: you publish an ai-catalog.json at a well-known path on your domain, registries index it, an agent searches by natural-language intent, then verifies the publisher and connects over MCP, A2A, or a plain API. It’s DNS plus a phone book for agents. The spec defines two mechanisms — a static manifest (ai-catalog.json) and a dynamic registry API (natural-language POST /search). A standard landing means a field to compete on just opened, the same way standardizing the Domain Name System created room for DNS providers and registrars. Right now the spec is still a draft, and the part that actually matters — a registry good enough to use, with strong index quality, ranking, and publisher verification — is still empty. The standard itself doesn’t decide who searches best. Which tool you rank first for a given intent is the product, and that seat is unclaimed.

How to Build It

Build a tool-discovery registry that implements the ARD spec. Three pieces matter. First, crawl and index — gather the ai-catalog.json files scattered across domains and index each tool’s description, representative queries, and tags as embeddings. Second, intent ranking — when an agent searches in natural language for “I want to do X,” return the right tools by confidence. Not keyword matching: weight publisher identity, compliance attestations, and real usage signals to raise precision. Third, a verification gate — don’t expose just any tool; verify publishers and filter out malicious or impersonating tools so the agent connects safely. Monetize on both sides: charge tool publishers (B2B fees for placement and priority indexing) and agent developers (per-call search-API pricing). Don’t try to cover every tool from day one; pick one vertical — say fintech connectors or data-pipeline tools — drive its index quality far above everyone else, and own that as the standard registry there. Reference implementations like Hugging Face’s Discover Tool already exist, so the differentiator is how much more precisely, and how much more safely, you find the right tool.