Classic AI code completion suggests the next line. Claude Code goes further: developers describe a task in natural language, and the tool plans the implementation, reads the affected files, changes code across multiple files, runs tests, and checks the result. This makes Claude Code an AI agent for software development, often also called a coding agent or AI programming assistant. In the background, it uses the language models of Claude.
According to Anthropic, Claude Code runs in the terminal, as an extension for VS Code and JetBrains IDEs, as a desktop app, and in the browser. Most variants require a Claude subscription or an Anthropic Console account.
How does Claude Code work?
Claude Code works in a loop of reading, planning, acting, and checking. Typical capabilities according to the documentation:
- Understand the codebase: Read files, recognize relationships between modules, and trace the causes of errors across multiple files.
- Change and run: Edit files, run commands, start tests, and fix errors.
- Work with Git: Stage changes, write commit messages, create branches, and open pull requests.
- Read project rules: A file named
CLAUDE.mdin the project contains guidelines such as coding standards, architecture decisions, or checklists. Claude Code reads it at the start of every session. - Connect tools: Via the Model Context Protocol, Claude Code accesses ticketing systems or documentation, for example.
- Automate: Hooks run custom commands before or after actions, such as formatting after every file change. In script mode, Claude Code can be used in pipelines, for example for reviews in GitHub Actions.
Permissions and security model
A tool that runs commands needs clear boundaries. Anthropic describes several layers for this:
- Permission modes: In manual mode, Claude Code starts with read-only permissions and asks before it changes files or runs commands. In auto mode, a separate classifier model checks the actions and blocks those it classifies as unsafe. Custom allow and deny rules apply in both cases.
- Sandbox: Shell commands can be run with file and network isolation.
- Working directory: In manual mode, Claude Code asks before its file tools read or write outside the project folder.
- Trust prompt: A confirmation appears on the first start in an unknown folder and for project-scoped MCP servers.
- Network commands: Commands that fetch content from the internet are not automatically allowed by default.
Anthropic explicitly points out that users must review suggested code and commands themselves before approving them and that no system is immune to all attacks. Anthropic does not security-audit MCP servers.
Typical use cases
- Write tests for previously untested code, run them, and fix errors
- Update dependencies, for example with Composer in PHP projects, and follow up on the consequences in the code
- Explain legacy code and modernize it step by step, for example before an upgrade due to PHP End of Life
- Resolve merge conflicts and draft release notes
- Have pull requests checked automatically in advance, before a human takes over the code review
What companies should keep in mind
- Contract and data: Use commercial plans or the API for company code. According to Anthropic, inputs and outputs of commercial products are not used for training by default.
- No secrets in the project: Credentials do not belong in the repository or in prompts. Whatever the tool can read can end up in the model's context.
- Isolation: For sensitive projects, use a sandbox or development containers and define permissions per project. Grant tokens and keys according to the principle of least privilege.
- Four-eyes principle: Changes go through a branch, a pull request, automated checks in the CI/CD pipeline, and human approval, never directly into production.
- Plan for manipulation: Issues, README files, or web pages can contain hidden instructions, see Prompt Injection. Do not pass unknown content to the tool unchecked.
- Manage centrally: Define managed settings for the entire organization and monitor usage, for example via the OpenTelemetry metrics supported by Anthropic.
How it differs
A chat assistant answers questions about code, whereas Claude Code works directly in the project and carries out steps. Compared with pure autocompletion in the IDE, it takes on entire tasks. In both cases, responsibility for the result remains with the development team.
How TOMORIS relates
The internal platform Werft shows how TOMORIS develops its own software with isolated working copies, tests, and reviewable pull requests. For your own project, you can get started under Custom Software.