honeydo

Honey, do everything. 🍯

One CLI for every AI capability — chat, vision, image, video, sound effects, TTS — cloud and local, built for AI agents.

npm i -g honeydo

honeydo ask "explain this code"              # LLM chat (claude / gemini / any OpenAI-compatible)
honeydo vision "how many ducks?" -i a.png    # vision understanding
honeydo image gen "a cat under sakura"       # local image gen (Apple Silicon)
honeydo image gen "logo" --engine doubao     # cloud image gen (Volcengine)
honeydo video gen "waves at dusk" --seconds 5
honeydo tts "你好世界"                        # MiniMax TTS
honeydo sfx gen "rain on a tin roof"
hd ask "hi"                                  # short alias, same thing

Named after the honey-do list — "honey, do this, honey, do that." You ask, it does. 🐝

Why a CLI, not MCP?

MCP servers inject their tool schemas into the agent's context on every single call — 10 tools you'll never use still cost tokens. A CLI is invoked only when needed, with zero standing context cost. honeydo was born from migrating five MCP servers (Gemini, Doubao, MiniMax, …) off MCP into plain commands.

Agent-friendly by contract:

• Unified exit codes: 0 ok / 1 runtime error / 2 bad args
• stdout/stderr discipline: results on stdout, progress/humans on stderr — pipes stay clean
• JSON output where it matters (--json on chat/vision, structured results on media gen)

Capabilities

Command	What	Engine
honeydo ask	LLM chat	Claude Code CLI / Gemini (agy) / direct Anthropic-compatible HTTP (--backend agy|api)
honeydo ask --backend local	chat against any local OpenAI-compatible endpoint	llama.cpp, vLLM, Ollama…
honeydo vision	image understanding	same local endpoint (Qwen-VL style)
honeydo image gen/edit/upscale/serve	image generation, editing, 2x upscale, daemon	Qwen-Image-2512 local (bf16, Apple Silicon)
honeydo image gen --engine doubao	cloud image gen	Volcengine Ark (seedream) with model fallback chain
honeydo video gen/setup	text-to-video, first-frame conditioning	mmh3turbo (MiniMax-H3 GGUF) local
honeydo sfx gen/batch/trim/normalize/...	full SFX production line	Dasheng-AudioGen local
honeydo lora list/add	LoRA registry (style/character/speed)	local Qwen-Image LoRAs
honeydo tts	text-to-speech, word-level subtitle timestamps	MiniMax speech-02
honeydo voice clone/list	voice cloning / voice inventory	MiniMax
honeydo doctor	local stack self-check	—

Install & configure

npm i -g honeydo

Cloud capabilities need API keys via env (nothing is stored by honeydo itself):

Env var	Used by
DOUBAO_API_KEY	image gen --engine doubao (Volcengine Ark)
MINIMAX_API_KEY	tts / voice (MINIMAX_API_HOST to override endpoint)
QWEN_API_URL / QWEN_MODEL / QWEN_API_KEY	ask --backend local / vision (default http://127.0.0.1:8001)

honeydo ask (claude backend) can optionally read providers from cc-switch's database when present — but works fine with a plain claude CLI login too.

Local media stack (Apple Silicon)

image / video / sfx run fully local via a Python inference stack (diffusers + torch MPS, ~110GB of model weights for the full image+edit setup). Run:

honeydo doctor    # prints what's ready, what's missing, and how to set it up

Security notes

• API keys are read from env vars only; honeydo never writes them anywhere.
• For the claude backend, the selected provider token is passed to the claude child process via argv (--settings) — visible in ps while the call runs. Prefer the api backend (pure HTTPS, no child process) if that matters to you.
• Local inference never leaves your machine.

Legacy bins

If you used the standalone tools, they still exist and remain compatible: gcli, lmedia, doubao, minimax. (qwen is deprecated — it collides with Alibaba's official qwen CLI; use honeydo vision / honeydo ask --backend local.)

Roadmap

See ROADMAP.md — unified config file, global --json, one-command local stack setup, i18n, brew tap.

中文文档：README.zh-CN.md

License

MIT