Vanor

The AI agent that runs at the speed of Node.js — zero deps, zero bloat, pure power.

Vanor is a CLI coding agent powered by LLMs. It reads your codebase, writes files, runs commands, and gets things done — all inside your terminal. No Python. No Docker. No 300MB node_modules. Just one install and you're flying.

万佑智算通用 CLI 智能体。零依赖，极速响应，终端即战场。📖 中文文档 →

Why Vanor

npm i -g @jpssff/vanor    # ← That's it. Nothing else.
vanor

• 0 dependencies. Not "few" — literally zero. Every line is vanilla Node.js. Installs in under a second. Starts instantly.
• Blazing fast. Native fetch streaming, native readline, native everything. No lag, no overhead, no nonsense.
• Battle-tested. Full Agent Loop with tool iteration, session persistence, memory, skills, and streaming LLM.
• Privacy-first. Everything stays local. Config, sessions, memory — all on your disk.
• Multi-model. OpenAI, Anthropic, or any compatible endpoint. Switch models mid-session with /model.
• Safe by default. Dangerous operations ask first. Allowlist, denylist, workspace boundaries. You stay in control.
• Skills that stick. Automatically loads skills from your workspace and other agents. Compatible with agentskills.io SKILL.md format.
• MCP-ready. Connect any Model Context Protocol server — filesystem, databases, APIs, custom tools. Hot-swap servers mid-session with /mcp.
• Speaks your language. Auto-detects Chinese or English. Switch anytime with /language.

One command to rule them all

vanor              # start chatting
vanor doctor       # health check
vanor resume       # pick up where you left off
vanor skills       # see what skills are loaded

Inside a session:

/auto-run          # trust me, go full speed
/compact           # reclaim context
/messages          # what did I ask again?
/skills            # what can you do?
/exit              # that's enough for today

Extend with MCP

/mcp add filesystem -- npx -y @modelcontextprotocol/server-filesystem /path/to/project
/mcp list           # see all connected servers
/mcp tools <name>   # explore what a server provides
/mcp disable <name> # temporarily turn one off
/mcp remove <name>  # clean up

Any MCP server's tools appear as mcp__<server>__<tool> — your agent can call them just like built-in tools.

It just works

$ cd my-project
$ vanor
▸ session sess_abc123 restored (3 messages)
▸ 49 skills loaded

You: refactor the auth module to use JWT

Vanor: [reads files → writes code → runs tests → reports results]
       Done. 4 files changed, all tests pass.

No setup wizard asking 12 questions. No YAML config hell. No "please install these 47 peer dependencies." Just your project and an agent that understands it.

What's inside

Agent Loop · Multi-model · Streaming · MCP · Tools · Skills · Memory · i18n · Privacy-first · Zero deps

The numbers

| | Vanor | Typical agent | |---|---|---| | Dependencies | 0 | 200–600+ | | Install size | ~300KB | 50–300MB | | Cold start | <100ms | 1–5s | | npm install | <1s | 10–60s | | Memory (idle) | ~30MB | 80–200MB |

Requirements

Node.js ≥ 18. That's it.

Philosophy

Vanor believes an AI agent should be a sharp tool, not a heavy platform. It does one thing — help you code — and does it with zero ceremony. No frameworks, no dependency chains, no supply chain risks. Just pure Node.js, clean architecture, and a relentless focus on speed.

License

MIT • Wanyou Intelligence

npm i -g @jpssff/vanor && vanor — your terminal just got smarter.