# Claude Code 开发者 Boris Cherny：用循环替代手动提示；neo-mcp 为 Claude Code 配备本地 AI 工程工人 Neo

- 来源：Rohan Paul (@rohanpaul_ai)
- 发布时间：2026-06-16 23:05
- AIHOT 分数：61
- AIHOT 链接：https://aihot.virxact.com/items/cmqgsrtas007fslpu03h6kfzx
- 原文链接：https://x.com/rohanpaul_ai/status/2066899865205178603

## AI 摘要

Boris Cherny（Claude Code 开发者）表示不再手动提示 Claude，而是编写循环让循环工作。@withneo 新发布的 MCP 服务器 neo-mcp 为 Claude Code 配备了本地 AI 工程工人 Neo。Claude Code 可将复杂 AI/ML 任务委托给 Neo，由其实现代码、运行实验、评估结果、调试失败运行并返回完整执行轨迹。在基准测试中，Claude Code + NEO 将任务成本从 $1.96 降至 $0.74，运行时间缩短 37%，并将后端从 PyTorch 切换为 ONNX Runtime 以优化 CPU 执行。该 MCP 服务器支持连接 Claude Code、Cursor、VS Code 等客户端，编辑器控制对话，Neo 负责执行，Claude 专注于决策而非执行噪音。

## 正文

"I don't prompt Claude anymore. I write loops and the loops do the work. My job is to write loops." - Boris Cherny， creator of Claude Code.

@withneo 's newly launched MCP server gives Claude Code a local AI engineering worker through neo-mcp.

It lets Claude Code hand off complex AI/ML tasks to Neo， which can implement code， run experiments， evaluate results， debug failed runs， and return the full execution trail.

I.e. Claude Code talks to you， Neo does the longer ML work.

In one benchmark， Claude Code + NEO cut task cost from $1.96 to $0.74， made runtime 37% faster， and switched the backend from PyTorch to ONNX Runtime for CPU-optimized execution.

Connect Claude Code， Cursor， VS Code， and other MCP clients to NEO. The editor stays in control of the conversation. NEO does the AI engineering.

So Neo becomes the execution layer for AI engineering work.

Claude Code hands off the task. Neo plans， runs experiments， monitors progress， evaluates results， and returns the trail： transcripts， files， metrics， reports， or repo changes.

That feedback loop matters.

Instead of burning Claude Code's limited iterations babysitting long-running tasks， Claude can use Neo's MCP tools to check status， pause， resume， inspect the trail， or ask for input at decision points.

Claude stays focused on decisions， not execution noise.