New on the blog: Wheels + Claude: Building a Feature via the stdio MCP 🔧 There are two ways to make an AI assistant useful inside a framework. The lazy one is a chat sidecar that generates snippets you paste. The good one is teaching the assistant the framework's own vocabulary -- and that's what Wheels 4.0 ships. The model doesn't write code; it decides which CLI command to run, and the framework generates everything exactly the way it always has. The whole MCP server is just

cli/lucli/Module.cfc

-- its public functions become tools by reflection. No port, no socket, no running dev server. The AI editor spawns

wheels mcp wheels

and speaks JSON-RPC 2.0 over stdio. What the post covers: • The reflection model: the CLI is the MCP server -- no second surface to drift • Setup with one

.mcp.json

snippet (no

wheels mcp setup

-- that command doesn't exist;

wheels mcp

just prints the config and writes nothing) • The 18-tool surface and exactly why

console

, `start`/`stop`,

new

, and friends are excluded • A worked example: shipping a commenting feature end-to-end through Claude Code -- generate, migrate, test • The deprecated

/wheels/mcp

HTTP endpoint and why stdio replaced it • A bit of template drift we found while writing the post (the OpenCode template still pointed at the old HTTP endpoint with a literal

{PORT}

) The tools advertise bare names --

generate

,

migrate

,

test

,

destroy

. Clients namespace them under the server, so Claude Code shows them as

mcp__wheels__generate

. <https://blog.wheels.dev/posts/wheels-claude-stdio-mcp>