CLI, MCP and library

The same 66 tools are exposed three ways. There is one implementation per tool; the surfaces are thin adapters over it.

CLI

cleya <tool> [options]

Results are JSON on stdout — exit code 0 on success, 1 on failure. Colour and the branded header appear only when stdout is an interactive terminal, so piping or redirecting gives you clean, parseable output. NO_COLOR is honoured.

cleya --help              # tools grouped by category
cleya <tool> --help       # options for one tool
cleya recipes             # goal → tool-sequence discovery
cleya capabilities        # the whole manifest, machine-readable

Flag grammar is uniform: --output is canonical with --out as a universal alias, underscores and hyphens are interchangeable, and comma-lists are equivalent to repeated flags.

MCP

{
  "mcpServers": {
    "cleya": { "command": "cleya-mcp" }
  }
}

Tools are registered as cleya_<tool> with Zod schemas. The server's instructions steer an agent toward the right path rather than making it guess — the single biggest failure mode in practice was never capability, it was discoverability: a capable agent hand-rolling FFmpeg or raw Chrome because it could not find the tool that already did the job.

Start any session with cleya_capabilities — one call returns every tool with its options, examples, category, cost hint and platform constraints.

Library

import { render, draft, contrast, brandCheck, resolveConfig } from "cleya";

const config = resolveConfig({ ffmpeg_path: "/opt/homebrew/bin/ffmpeg" });
const result = await draft({ starter: "promo-hero", brand: "brand.json" }, config);

Every tool is (params, config) => Promise<ToolResult<T>>, fully typed. Core helpers are exported too — resolveConfig, the platform presets, the scene compiler, the starter loader and the harvest primitives.