Local coding agent

查看原文

Claude模型发布开源代码端侧推理

热度趋势

趋势数据积累中

百分比基于当前可用热度信号，而非评论数或独立用户人数。

推荐理由

Claude 相关模型动态已经出现，适合跟踪能力变化、生态影响和后续可用性。

AI 摘要

一位开发者发现使用Claude进行编码的成本过高，因此开始寻找替代方案。他们首先尝试了Cline，但效果不佳。随后，他们转而使用VS Code中的Copilot聊天功能，并将模型设置为“Auto”，结果发现它在代码转换方面表现出色。现在，这位开发者正在向社区征求意见，希望了解其他人使用的编码代理和工作流程。

I’ve been using Claude for coding for quite a while, but it eventually became a bit too expensive for the amount I was using it.

I recently got myself a 5090 and thought I’d give local models a serious try. I already have VS Code set up, but I’m still trying to figure out what the “right” workflow is.

One of the first things I wanted to do was convert a small application I have written in PowerShell over to Python, while keeping the existing GUI and functionality as close to the original as possible.

I tried Cline first, but it didn't really manage to get there. Then I used the Copilot chat inside VS Code with the model left on Auto, and surprisingly it handled the conversion pretty easily.

That made me curious, so I tried switching Copilot from Auto to Qwen3-Coder. Interestingly, Qwen struggled with the exact same task where Auto had done a pretty good job.

Now I’m wondering how much of this comes down to the actual model and how much comes down to the agent/environment around it.

For those of you using a 5090 or similar hardware for local coding, what has actually worked well for you?

I’m particularly interested in setups where you can throw an existing project at the AI and let it understand the codebase, make changes across multiple files, run things, debug, etc., rather than just asking it to generate individual pieces of code.

I’m not really interested in benchmark results — I’m more interested in what you’ve found reliable in actual day-to-day development.

Would love to hear what you’re using and how your setup/workflow looks.