Exploring Claude Code: How Anthropic’s AI Agent Transforms Coding

Anthropic recently released a beta research preview of Claude Code, an AI‑powered coding agent that leverages the Claude 3.7 Sonnet model to automatically perform tasks such as adding features, testing, creating pull requests, and generating documentation.

Claude Code uses a command‑line interface (CLI) as its user interface and currently has no IDE or editor plugins.

Anthropic engineer Erik Schluntz explains that because Claude Code runs in the terminal, developers can bring it into any IDE or server they prefer.

He adds that the tool is popular internally at Anthropic, with many engineers pairing it with IDEs such as Cursor, Windsurf, VS Code, Zed, and Xcode; a typical workflow starts in Claude Code for most coding tasks and finishes in an IDE.

The AI agent runs on Node.js 18 or higher, can interact with Git, GitHub, and GitLab, and cannot run on Windows except via the Windows Subsystem for Linux. It works in environments like the VS Code terminal, although documentation warns that the VS Code terminal may truncate long pastes.

Common tasks Claude Code can handle include explaining code, automating Git operations, refactoring, fixing bugs, and writing tests and documentation.

In a demo video the team used Claude Code to add a chat‑history feature to a Next.js application.

Claude Code adds functionality to a Next.js app example

A developer noted that while the example targets React, Claude would happily ignore that context and produce Svelte code if asked.

Claude provides a set of tools such as BashTool for executing shell commands and GrepTool for searching patterns in file contents. Configuration lets users specify which tools may run without manual approval, which files or directories to ignore, and notifications upon task completion.

Aside from complaints about API rate limits and cost, early reactions to the research preview have been positive.

One developer said, “I was refactoring a complex legacy React component, installed Claude Code, and it completed the job quickly. I like the interface and its personality.”

For security reasons, developers can use the flag --dangerously-skip-permissions (e.g., in an offline Docker container) to bypass manual approvals, but running the AI agent unsupervised carries risk because it may make mistakes; AI agents are more powerful than simple code‑editing AIs, so the risk is higher.

Cost is a concern: “The tool consumes a lot of tokens,” said an early adopter. Cost tracking is simple—Claude Code prints the cost each time a session ends, and users can run /cost during a session to see accumulated spend, with the ability to set spending limits.

Users are reminded that this is a preview version and may encounter issues during development.

Anthropic technical staff Boris Cherny clarified that Claude Code is a research preview and therefore “rough around the edges,” unlike Cline, which is a polished VS Code AI assistant that works with various models.

Regarding privacy, Anthropic promises not to use feedback obtained from Claude Code to train its models, though users can opt‑in to share usage data for debugging and product improvement.

Anthropic views Claude Code as a step from AI‑assisted coding toward “expert‑level” collaboration, even though some observers warn that AI could degrade code quality.

Inputting “claude commit” lets the AI generate commit messages and execute Git commands, accelerating developers’ workflows and likely becoming commonplace soon.

编辑：聆听音乐的羊 相关参考： https://www.anthropic.com/news/claude-3-7-sonnet https://davidjarman.net/archive/2025/02/24/claude-code-initial-impressions

相关阅读：

在 Visual Studio Code 中免费使用 DeepSeek R1

初学者实际上已经无法再进行编码

2025 年编码将会消亡吗？