nonatofabio/luna-agent

月球特工

一个自定义的最小AI代理，具备持久内存、MCP工具集成、Discord/CLI接口和结构化可观察性。 完全运行在本地硬件上——无需云API费用。

---

~2300行Python。没有框架。

为什么是习俗

我们评估了现有的代理框架，并全部拒绝了：

• OpenClaw：40万行代码，4.2万个暴露实例在Shodan上。太大无法审计，太大无法信任。
• 零爪：评估时9天大。太幼稚了。
• 纳米爪：太薄了——反正大部分都得重建。

核心需求（内存、工具、聊天、日志）都是单独解决的问题。不需要40万线的框架。

建筑

所有LLM流量都经过一个带有可配置端点URL的单一LLMClient。如今它指向localhost:8001（llama-server）。若要后续插入AI防火墙，只需将URL改为localhost:9000——无需任何代码更改。

思考模型支持： Luna 自动处理推理模型（Qwen3.5 等）——提取 reasoning_content，内容空时回退到清理推理，并从输出中剔除泄漏的标记（ 、 等）。

硬件

Luna Agent

A custom minimal AI agent with persistent memory, MCP tool integration, Discord/CLI interface, and structured observability. Runs entirely on local hardware — no cloud API costs.

---

~2300 lines of Python. No frameworks.

Why Custom

We evaluated existing agent frameworks and rejected them all:

• OpenClaw: 400K lines of code, 42K exposed instances on Shodan. Too large to audit, too large to trust.
• ZeroClaw: 9 days old at time of evaluation. Too immature.
• NanoClaw: Too thin — would need to rebuild most of it anyway.

The core needs (memory, tools, chat, logging) are individually well-solved problems. No 400K-line framework needed.

Architecture

Discord (discord.py)     CLI REPL (no token)
     |                        |
     v                        v
+---------------------------------+
|        Luna Agent Core          |
|                                 |
|  agent.py                       |  agent loop: msg → memory → prompt → LLM → tools → respond
|    ├── llm.py                   |  single LLM client, configurable endpoint
|    ├── memory.py                |  SQLite + FTS5 + sqlite-vec hybrid search
|    ├── tools.py                 |  native tools: bash, files, web, delegate, code_task
|    ├── tool_output.py           |  smart output pipeline for large results
|    ├── mcp_manager.py           |  MCP client for community tool servers
|    └── observe.py               |  structured JSON logging
|                                 |
+---------------------------------+
              |
              v
        llama-server               Qwen3.5-35B-A3B on 2x RTX 3090

All LLM traffic flows through a single LLMClient with a configurable endpoint URL. Today it points at localhost:8001 (llama-server). To insert an AI firewall later, change the URL to localhost:9000 — zero code changes required.

Thinking model support: Luna handles reasoning models (Qwen3.5, etc.) automatically — extracting reasoning_content, falling back to cleaned reasoning when content is empty, and stripping leaked markup (, , etc.) from output.

Hardware

• Intel i7-13700K, 64GB DDR4
• 2x NVIDIA RTX 3090 (24GB each, 48GB total)
• Qwen3.5-35B-A3B Q8_0 via llama-server with layer split across both GPUs
• 131K context window, Q8_0 KV cache

Quick Start