Anthropic has released a “beta research preview” of Claude Code, an agentic coding tool which uses the Claude 3.7 Sonnet model to automate tasks including adding features and tests, creating pull requests, and generating documentation.

Claude Code uses the command line for its user interface, rather than integrating with an IDE or editor via a plugin. Anthropic engineer Erik Schluntz explained that “because Claude Code is just in the terminal, you can bring it into any IDE (or server) you want.”

He added that “Claude Code is super popular internally at Anthropic. Most engineers like to use it together with an IDE like Cursor, Windsurf, VS Code, Zed, Xcode, etc. Personally I usually start most coding tasks in Code, then move to an IDE for finishing touches.”

The agent runs on Node.js 18 or higher and interacts with Git, GitHub and GitLab. It does not run on Windows other than via Windows Subsystem for Linux. It works in environment such as the Visual Studio Code (VS Code) terminal, though the documentation warns that the VS Code terminal is “particularly prone to truncating long pastes.”

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

The tool can also add features, and in an example video the team used Claude Code to add a chat history feature to a Next.js application.

This example is optimal for Claude since it does seem to be oriented towards React. “I might be working on a Svelte codebase and Claude will happily ignore that context and provide React code,” a developer complained.

Claude code has a set of tools such as BashTool which executes shell commands and GrepTool which searches for patterns in file contents. Configuration includes options for which tools can run without manual approval, files and directories to ignore, and notifications when tasks complete.

Early reaction to the research beta is positive, other than complaints about API rate limits and cost. “I’m in the middle of a particularly nasty refactor of some legacy React component code … I decided to install Claude Code and it is making short work of this. I love the interface. I love the personality,” said one comment.

There are potentially a lot of manual approvals, for safety reasons, though coders can use the flag –dangerously-skip-permissions, for example in a Docker container without internet, to avoid these if desired. Otherwise it is risky to allow an AI agent to work without supervision since it may make mistakes, and agentic AI is more powerful and therefore more risky than AI that only edits code.

The cost is an issue. “This tool uses A LOT of tokens right now. I mean, it’s breathtaking,” said another early adopter.

Tracking spend is straightforward. “Every time you exit a session we print the cost, and in the middle of a session you can do /cost to see your cost so far that session,” explained Schluntz. It is also possible to set spend limits.

This is a preview, and developers should expect issues. “Claude Code is a research preview — it’s more rough, lets you see model errors directly, etc. so it’s not as polished as something like Cline,” said Anthropic technical staff member Boris Cherny. Cline is an add-in AI assistant for VS Code that can work with a variety of models.

Regarding privacy, Anthropic promises not to “train generative models using your feedback from Claude Code” though feedback including transcripts of usage may optionally be used for debugging and product improvement.

Anthropic sees Claude Code as a step from AI assistance to collaboration “on par with experts” – though some others find that AI can be detrimental to code quality.

There is no doubt though that typing “claude commit”, for example, and getting the AI to write the commit message and execute the Git command, speeds developer workflow and will likely become something that is taken for granted before long.