MCP Overview

dlazy exposes the entire tool registry through a single Streamable HTTP MCP endpoint. Authenticate with a Bearer key, optionally narrow the surface with query params, and you're done.

Endpoints

Cloud: https://dlazy.com/mcp
Local dev: http://localhost:3000/mcp (when running this repo with pnpm run dev)

Authentication

Every request must carry the Bearer header:

POST https://dlazy.com/mcp
Authorization: Bearer sk-xxxx

Query Parameters

Param	Effect
`pid`	Bind generations to a specific project
`models`	Comma-separated `cli_name`s — narrows the exposed tool set
`generate=true`	Enable auto-generation hints (server decides eagerly)

Tool Naming

The MCP server (lib/ai/mcp.ts) registers each tool by its bare cli_name (e.g. gpt-image-2, veo-3.1). Clients prefix the name however they choose:

Client	Surface name
Claude Code	`mcp__dlazy__gpt-image-2` (per Claude Code's `mcp__<server>__<tool>` convention)
OpenClaw / generic MCP	`gpt-image-2` (raw)

Async Tools

Async tools (most video models) wait by default on the MCP side — the client sees one synchronous reply per call. Use the CLI's --no-wait + dlazy status if you need detached behaviour.