Claude Code brings Anthropic's coding assistance into development workflows.
Claude Code
Explore features, practical uses and pricing below.
Claude Code is Anthropic's coding agent for working with a repository rather than only discussing a pasted code snippet. It can read files, make changes, run development commands and connect to tools. Supported surfaces include the terminal, editor integrations, desktop application and browser. It suits developers who can explain a task, provide project context and assess the resulting code and command output.
The workflow documentation covers exploring unfamiliar code, tracing a bug, refactoring, writing tests and preparing pull requests. A request can span several files: for example, tracing a login problem from the interface to an API and database layer. Plan mode provides a way to investigate and review an approach before file changes begin. Separate worktree sessions can help organize concurrent work without placing every edit in the same checkout.
The overview explains the interface differences. The VS Code extension provides inline diffs, file mentions and plan review. The desktop application supports visual diff review and multiple sessions, while the terminal CLI fits workflows that already use shell tools. Choose the surface that makes it practical to inspect the agent's work; the interfaces do not all have identical setup or access requirements.
Project instruction files can record build commands, conventions and architectural context that would otherwise need to be repeated. Claude Code also has auto memory for notes based on corrections and preferences. The documentation distinguishes these sources of context from enforced configuration: an instruction saying to avoid an action is not the same as a permission rule that blocks it.
Permission controls govern file access, edits, shell commands and other tools. Modes and managed policies affect which actions require approval. Developers should review the active configuration for the repository and chosen surface, especially before allowing commands or connecting external tools that can change systems beyond the project.
Give Claude Code a reproducible failure in an order-total calculation, point it to the relevant module and explain the expected rounding rule. First request an investigation and proposed change. After reviewing the plan, let it implement a focused patch and run the relevant tests. Inspect the diff and test assertions yourself, including a boundary case the original implementation mishandled. Ask for an explanation of unresolved failures before deciding whether to commit the result.
A repository-reading agent can still infer the wrong requirement, modify unrelated code or produce tests that merely confirm its own implementation. Clear acceptance criteria and small changes make these problems easier to detect. Long sessions also accumulate context, which can affect token usage and how earlier information is retained. Use version control and review the actual command output rather than relying only on the agent's completion message.
Local tools, integrations and cloud sessions have different environments. A successful command in one session does not establish that a deployment or another machine has the same dependencies and permissions.
Access depends on the surface and account arrangement. The overview describes Claude subscriptions, Anthropic Console accounts and supported third-party provider options for selected integrations. The usage and cost guide distinguishes subscription allowances from API token billing and provides usage tracking. Confirm eligibility, model usage limits and any extra usage charges for the intended account. A local session-cost estimate is not necessarily the amount charged to a subscription.