Claude Code is Anthropic's AI assistant for programming. You describe in plain language what you want, such as "fix the error on the checkout page", and it reads the project's code, edits files, runs tests and commands, and can prepare the change for review. For a business with a website or webshop, it lets a developer hand over routine work such as writing tests, fixing bugs and updating dependencies.
When did it appear?
Anthropic introduced Claude Code on 24 February 2025 as a command-line tool in a limited research preview, together with Claude 3.7 Sonnet (announcement). It became generally available on 22 May 2025, with extensions for VS Code and JetBrains and background tasks through GitHub Actions (Introducing Claude 4). In March 2026 its source code leaked through a source map in an npm release (what the leaked source revealed). In September 2026 the documentation lists the terminal, VS Code, JetBrains IDEs, a desktop app and the browser, and most of these need a Claude subscription or an Anthropic Console account.
How does it work?
Anthropic calls Claude Code the harness around a Claude model: the layer that supplies the tools and manages what the model sees (How Claude Code works). A task runs as an agent loop of three phases: gather context, take action, verify results. Asked to fix failing tests, it might run the test suite, read the errors, find and edit the relevant files and run the tests again. You can interrupt it at any point.
At the start of every session it reads CLAUDE.md, a Markdown file with the project's instructions, and it keeps an auto memory of what it learned. When the context window fills up, it clears old tool output and then summarises the conversation, so early instructions can get lost. The documentation advises keeping lasting rules in CLAUDE.md.
A permission mode sets what it may do without asking. In manual mode it asks before every file edit and shell command, and in plan mode it only explores and proposes a plan. In September 2026 auto mode, in which a classifier blocks risky actions, is the starting mode in the terminal and VS Code on Pro, Max and Team plans. Before each edit it saves a checkpoint, so file changes can be undone; changes to databases, APIs or deployments cannot.
How can you extend it?
- MCP connects it to outside tools such as Jira, Slack or Google Drive. With
claude mcp serveit can also act as an MCP server for other applications (MCP docs). - Skills package repeatable workflows, such as a
/review-prcommand. Hooks run shell commands before or after its actions, such as a formatter after every edit. - Subagents take on a subtask in their own context window and report back a summary.
- Plugins bundle skills, subagents, hooks and MCP servers for sharing, such as Jesse Vincent's Superpowers, which had 14 skills for planning, testing and code review in April 2026.
What are routines?
Since 14 April 2026, routines let Claude Code work while your computer is off. A routine is a saved prompt with one or more repositories and connectors, started on a schedule, by an API call or by a GitHub event such as a new pull request. It runs on Anthropic's cloud and pushes its changes to branches whose names start with claude/. A routine does not stop to ask for permission, so the routines documentation advises limiting its repositories, network access and connectors to what it needs. In September 2026 routines are still a research preview, on Pro, Max, Team and Enterprise plans.
Example
Suppose the developer of a small webshop types in Claude Code: "Write tests for the discount-code module, run them and fix any failures." Claude Code finds the module, writes the tests and runs them. If some fail, it reads the errors, adjusts the code and runs the tests again, and in manual mode it asks before each command. The developer reviews the change and asks Claude Code to commit it.