AI agent discovery: MCP, OpenAPI, and machine-readable capability catalogs
How AI agents discover tools and APIs using MCP, OpenAPI, manifests, agents.json, llms.txt, and live capability catalogs.
Short answer
AI-agent discovery works best when providers publish structured descriptions that software can inspect directly. MCP exposes callable tools to compatible clients; OpenAPI describes HTTP interfaces; manifests and agent catalogs provide stable entry points; and LLM context files give models concise orientation.
MCP vs OpenAPI
| Surface | Best use |
|---|---|
| MCP | Tool discovery and invocation from MCP-capable AI clients. |
| OpenAPI | HTTP endpoint schemas, operations, request parameters, and response structures. |
| agents.json | Provider-specific machine-readable catalog and discovery metadata. |
| llms.txt | Concise model-facing orientation and links to authoritative machine surfaces. |
| Well-known manifests | Predictable locations that reduce discovery ambiguity. |
Why discovery matters
An agent cannot reliably choose a capability it cannot understand. Discovery metadata should answer: What does this service do? Where is its live contract? How is it invoked? Is payment required? What is the health state? Which document is authoritative for the current capability inventory?
A practical discovery chain
1. Start at a stable corporate manifest. 2. Follow links to live service catalogs. 3. Inspect the operation/tool needed for the task. 4. Evaluate price and availability. 5. Invoke through the supported protocol.
Script Master Labs discovery entry points
https://www.scriptmasterlabs.com/.well-known/mcp.json
https://www.scriptmasterlabs.com/.well-known/openapi.json
https://www.scriptmasterlabs.com/agents.json
https://www.scriptmasterlabs.com/llms.txt
https://mcp-x402.onrender.com/mcp
Do not treat hard-coded capability counts as permanent truth
Tool and endpoint inventories change. A strong discovery layer points clients to the live contract rather than forcing crawlers to trust a stale number copied into marketing text.