Capabilities — Exposing Tools an Agent Can Call

The agent-readiness dimension that exposes tools an agent can call — a JSON API, MCP server, WebMCP and agents.json — so agents act, not just read.

What does the capabilities dimension mean?

The capabilities dimension turns a site an agent reads into a site an agent can use: typed, declared actions — search a catalog, check a booking, submit a form, run a query — exposed through a machine-callable interface instead of the human UI. It is optional and sits late in the implementation order: implement it only when agents should act on your behalf.

Which signals and standards do capabilities cover?

Four capability surfaces, from a plain API to in-browser tools:

• JSON API — a documented endpoint that returns structured data, the simplest capability surface. RFC 9727 (IETF, 2025) gives it a discovery point: /.well-known/api-catalog, a Linkset that lists a publisher's APIs.
• MCP (Model Context Protocol) — open-sourced by Anthropic on 25 November 2024 (Anthropic, 2024): a JSON-RPC 2.0 protocol in which servers expose tools, resources and prompts, hosted since 9 December 2025 by the Agentic AI Foundation under the Linux Foundation (Linux Foundation, 2025). Its current protocol revision is dated 2026-07-28 (Model Context Protocol, 2026).
• WebMCP — a W3C Web Machine Learning Community Group draft, edited at Microsoft and Google and dated 4 September 2026, that is not a W3C Standard; it lets a page register JavaScript tools on document.modelContext through registerTool(), so an agent driving the browser can call them (W3C WebML CG, 2026).
• agents.json (action manifest) — the /.well-known/agents.json manifest that declares which interfaces and actions a site exposes; Wildcard AI's spec builds it on OpenAPI (Wildcard AI, 2025). Its location is the discoverability facet of the same file.

How does the Agent-Readiness Audit score capabilities?

The Capability category holds 3 of the Audit's 18 checks and 17 of 100 points: agents-manifest (7) passes when /.well-known/agents.json returns 200 and parses as JSON; json-api (5) and webmcp (5) pass when that manifest advertises interfaces.json_api and interfaces.webmcp (Agents Welcome, 2026). The Audit verifies the declaration, not the endpoint's behavior. This site proves it live: data/tools.json is the single source for its 12 WebMCP tools, served at /api/tool-registry, registered on the homepage's document.modelContext and mirrored into agents.json as interfaces.webmcp (count 12); every tool has an HTTP twin, so an agent without JavaScript gets the same result.

Capabilities — frequently asked questions

Do I need MCP to be agent-ready?

No. Capabilities are optional; discoverability, content and access control come first. A documented JSON API listed in /.well-known/api-catalog is the simplest callable surface, and the Audit awards its 17 capability points for a valid agents.json that advertises a JSON API and callable tools.

What is the difference between MCP and WebMCP?

MCP is a JSON-RPC 2.0 protocol between an AI application and a server you host; WebMCP is a browser API draft that lets the web page itself register tools on document.modelContext for an agent driving the browser.

How many callable tools does this site expose?

12, defined once in data/tools.json and served at /api/tool-registry, registered on the homepage's document.modelContext and listed in /.well-known/agents.json. Each has an HTTP twin, so no JavaScript is required for the same result.

Related: the agents.json spec · publish agents.json · the Agent Protocol Atlas · frontier models that complete tasks · audit your site

← Agent-Readiness · .md