Claude Code asks permission before it changes files or runs commands

A research preview dated 24 February 2025

Anthropic announced Claude Code on 24 February 2025, alongside a model release, describing it as 'our first agentic coding tool' and 'an active collaborator that can search and read code, edit files, write and run tests, commit and push code to GitHub, and use command line tools'. The announcement called it a 'limited research preview' and said only that Anthropic planned to 'continually improve it based on our usage' in the weeks after, without giving a further date. A cited internal figure — that it 'completed tasks in a single pass that would normally take 45+ minutes of manual work' — is Anthropic's own account of its own tool and is not an independently measured result.

What the current documentation says it does

Anthropic's current documentation and product page are living pages, describing the tool as it exists on 16 September 2026 rather than at its 2025 launch; both note recent feature additions dated well into 2026. As retrieved, the tool is described as reading a codebase, editing files, running commands and 'integrating with your development tools' across a terminal, an IDE extension, a desktop app and a browser. The product page states directly that it 'asks for permission before making changes to your files or running commands, and runs locally rather than requiring remote servers'. The documentation separately describes a VS Code extension offering 'inline diffs' and 'plan review', and a desktop app letting a developer 'review diffs visually' — a person is shown the proposed change, in a specific form, before or as it lands.

What stays the developer's job

None of this removes ordinary responsibility. A permission prompt only works if a developer reads it rather than approving by habit, and the documentation itself shows a tool still being reshaped: features named in the current pages, such as scheduled 'routines' or agents that run in parallel and merge each other's work, are recent additions, not the tool the 2025 announcement described. A team still decides what CLAUDE.md instructions, if any, constrain the tool's behaviour, still owns its git history and commit review once a change is pushed, and still has to configure whether permission is asked per action, granted for a session, or bypassed — a choice the documentation places in the developer's settings, not the model's judgement. Reading a diff quickly is not the same as reviewing it.

• Is permission requested before every file edit and command, or only for actions the tool classifies as risky?
• Who reviews a diff once it is created, and does that differ from the team's normal code review?
• Which claims about speed or capability came from the vendor's own account rather than an independent measurement?

A tool that asks first is a meaningfully different design from one that edits and reports back afterwards. It is still a design, described by the company that built it, and a team adopting it inherits the job of checking that the permission and review steps it relies on are actually being used.