wired for robots

Agents & MCP

A read-only MCP server on this site. Agents resolve a component the same way a person would: define the brand direction, compare candidates, install with the CLI, then adapt the source.

Protocol

  • Connect your MCP client to the URL below using Streamable HTTP. It supports initialization, tool discovery, and tool calls over JSON-RPC. No API key is required.
  • GET /mcp returns { name: "JabKit MCP", tools: string[] }.
  • POST /mcp also accepts the original HTTP API body { tool, arguments } and returns JSON.
  • The tool reference pages show direct HTTP examples. MCP clients discover the same tools and their argument schemas automatically.
https://jabkit.joseadrianbuctuanon.dev/mcp

Data source

Every handler reads the committed registry JSON under /r. The endpoint is exactly as fresh as the last registry build. There is no authentication and no rate limiting.

Agent flow

  1. 01get_conventionsAnswer JabKit's two discovery questions: what the page must accomplish and what it should look and feel like. Reuse the answers in brief on every call.
  2. 02search_componentsCompare candidates using tone, bestFor, avoidFor, visualWeight, layout, and capabilities. Inspect shortlisted source with get_component.
  3. 03get_install_planSee files, npm deps, and registry deps (or get_component for the full document).
  4. 04npx @jabkit/cli addInstall pristine source, then adapt composition, typography, brand tokens, and content. Replace demo branding and review desktop/mobile and supported themes.

Every tool returns needs_brief until the agent supplies both discovery answers. The agent answers from the existing conversation and asks the user only for missing details that affect selection. Calls with a complete brief return catalogue data.

Limitations

  • Read-only by design. The CLI is the writer.
  • Stateless requests; no persistent sessions or notification stream.
  • search_components returns eight results by default, with limit (up to 50) and offset; list_components has no pagination.
  • Search scores use lexical metadata matches. They do not judge visual quality or detect every semantic conflict; inspect source and previews.
  • filesToOverwrite is always empty.

Tools