What is Claude Code?
Claude Code is Anthropic's agentic coding product for inspecting a project, proposing and applying edits, running tools and commands, and working through development tasks. It is available through terminal, VS Code, JetBrains, desktop and web surfaces, but execution location, authentication, plans and feature availability differ. It does not place every repository file into every model request, run the model locally by default, or replace tests, code review and human merge ownership.
Why — the first-principles explanation
Claude Code is a coding product built around Claude models, project context and tools; it is not a separate model and not an autonomous software owner. A user supplies a goal, Claude selects or requests context, proposes actions, and Claude Code mediates file reads, edits, shell commands and connected tools. The useful unit is therefore a reviewable change that passes the project's acceptance gates—not a confident answer or a large diff.
Context is selected, not magically complete. Each session starts with a fresh context window. `CLAUDE.md` files can load project instructions, while Claude can search and read files in the working directory and any additional directories the user grants. Relevant file contents, tool results, conversation history and instructions compete for a finite context window and may be compacted. Claude Code can reason across many files, but “understands the codebase” should not be read as a guarantee that every repository file, generated artifact, dependency, secret or runtime state is present in every model request.
The action loop has two control layers. Permissions decide which tools and paths Claude may use and when approval is required. The optional Bash sandbox adds operating-system-level filesystem and network boundaries for shell commands and their child processes; it does not sandbox every Claude Code tool. `CLAUDE.md` can steer behavior, but Anthropic documents it as context rather than enforcement. Hard restrictions belong in permissions, sandbox policy, managed settings or hooks. MCP can add external tools and data, which also expands the trust, credential and prompt-injection boundary.
“Runs in the terminal” does not mean “runs the model locally.” In a local terminal or IDE session, the Claude Code client and approved tools operate on the user's machine, while prompts, selected code context and model output travel over the network to the active model provider. Claude Code on the web instead runs the task in an Anthropic-managed virtual machine and obtains repository access through its supported connection flow. Desktop can start local or cloud-oriented workflows. The surfaces share the Claude Code product layer, but session location, credentials, controls and available features are not interchangeable.
Access is also not one fixed subscription. The terminal supports Claude subscriptions, Anthropic Console access and supported enterprise providers including Amazon Bedrock, Google Cloud's Agent Platform and Microsoft Foundry. Anthropic's feature matrix says the CLI, IDE extensions, `CLAUDE.md`, MCP and local tooling work across providers, while web, mobile, Desktop, Remote Control and managed GitHub Code Review have subscription or plan boundaries. Managed GitHub Code Review is listed for Team and Enterprise; the local `/code-review` command is a separate workflow. Prices, usage limits, seats and availability change, so a static “starts at” price is not a safe definition.
Data handling depends on the account, provider and surface. Anthropic distinguishes consumer data controls from commercial terms, documents local plaintext transcript caching by default, and describes separate data flows for local and cloud execution. Before using private code, verify the active login and provider, training preference, retention, telemetry, repository connection, MCP servers and cloud environment. Claude can run tests and produce review findings, but those are evidence—not acceptance. The repository owner remains responsible for inspecting the diff, validating behavior and security, controlling credentials, and deciding whether to merge or deploy.
An example that makes it click
Suppose an authentication test fails after a dependency update. Start Claude Code on a clean branch in the repository, state the failing command and acceptance criteria, and ask it to identify only the files and runtime evidence needed. Review its plan before granting edits. Keep production secrets unavailable, approve the narrow commands required, and use permissions plus the Bash sandbox where supported. After the patch, rerun the reproduction, unit and integration tests, lint, type checks and a security scan; then inspect the diff and dependency changes yourself. If the same task is sent to Claude Code on the web, first verify the GitHub connection, cloud environment, network policy and data terms because execution moves to an Anthropic-managed VM. The result is acceptable only when the human owner confirms the evidence—not when Claude says the bug is fixed.
How to do it
- Choose the surface and execution location first: local terminal or IDE, Desktop, or an Anthropic-managed web session. Do not treat them as identical data or credential paths.
- Confirm the active authentication method with the current account settings: Claude subscription, Anthropic Console, Amazon Bedrock, Google Cloud's Agent Platform, Microsoft Foundry or an organization gateway.
- Classify the repository and check the current training, retention, telemetry, cloud-session and provider terms before exposing private or regulated code.
- Start from the correct repository, branch and working directory. Record a reproducible failure, expected behavior and the tests that decide acceptance.
- Keep `CLAUDE.md` concise and specific for build commands, architecture and coding conventions, but enforce hard restrictions through permissions, sandbox policy, managed settings or hooks.
- Ask Claude to explain its proposed scope, relevant files, assumptions and plan before a multi-file change. Never infer that every file or runtime dependency is already in context.
- Grant least privilege. Use explicit allow, ask and deny rules, enable the Bash sandbox where supported, and review any request to run outside its boundary.
- Add only trusted MCP servers, review their tools and credentials, and remember that an external tool can read data or cause side effects beyond the repository.
- Inspect the final diff and run independent tests, lint, type checks, security checks and any domain review required for production. A generated test that passes is not independent proof by itself.
- Compare live plans and providers on accepted changes, human correction time, latency, metered usage, data controls and failure recovery; re-check after material product or policy changes.
Key facts
- Claude Code is Anthropic's agentic coding product around Claude models, context, tools and approval controls; it is not a standalone model or an automatic code owner.
- Anthropic documents terminal, VS Code, JetBrains, Desktop and web surfaces, but the execution location, session behavior and feature availability differ.
- The terminal supports Claude subscription and Console accounts plus supported providers such as Amazon Bedrock, Google Cloud's Agent Platform and Microsoft Foundry; provider choice changes credentials, billing, data flow and some features.
- A local CLI session can read files and run approved commands on the user's machine, but model inference is networked: prompts, selected context and outputs go to the active provider.
- Claude Code can search and inspect multiple project files, but no official guarantee says every repository file or runtime dependency appears in every model request.
- Permissions and sandboxing are complementary: permissions govern tools and access, while the sandbox provides OS-level filesystem and network enforcement for Bash commands and their child processes.
- `CLAUDE.md` files carry project instructions into context but are not an enforcement layer; MCP servers add external tools and data and must be permissioned and trusted separately.
- Claude Code on the web runs on Anthropic-managed cloud infrastructure and accesses a repository through supported GitHub or upload flows rather than executing on the user's local machine.
- Anthropic separates consumer and commercial data policies and documents local session transcripts stored in plaintext for a configurable period by default; verify the current account, provider and retention settings.
- Managed GitHub Code Review is listed for Team and Enterprise, while local diff review is a separate CLI workflow. Neither replaces tests, human review or merge ownership.
Open the official product site and confirm the current access, plans and terms.
Decide whether Claude Code fits your repository workflow
Compare context selection, execution location, provider, permissions, data terms and cost per accepted change on the same repository tasks; keep tests and merge approval independent.
▶ The 60-second explainer (script)
What is Claude Code? It is Anthropic's agentic coding product, not a separate model and not an automatic software owner. It can inspect project files, propose edits, run approved commands and use connected tools through terminal, IDE, Desktop and web surfaces. But three boundaries matter. First, context is selected: Claude can search across a repository, yet every file and runtime dependency is not guaranteed to be in every model request. Second, a local terminal session is not local inference. The client and approved tools run on your machine, while prompts, selected code context and outputs go over the network to the active provider. Web tasks instead run in Anthropic-managed virtual machines. Third, instructions are not enforcement. `CLAUDE.md` guides Claude; permissions, sandbox policy and hooks set harder boundaries, and MCP servers add their own trust surface. Plans and features differ by login and provider, so check the live matrix instead of relying on a fixed price. Always inspect the diff, rerun independent tests and security checks, and keep merge and deployment approval with a human owner.
What authoritative sources say
People also ask
What is Claude Code?
Claude Code is Anthropic's agentic coding product. It combines Claude models with selected project context, file and shell tools, MCP integrations, permissions and review surfaces to help inspect, change and test software.
Is Claude Code an AI model?
No. Claude Code is the product and agent harness around Claude models. The active model and provider can vary, while the client supplies project context, tools, permissions and interfaces.
Does Claude Code read the entire codebase?
It can search and read many files in approved directories, but that is not a guarantee that every file, dependency, secret or runtime state is loaded into every model request. Ask what it inspected and verify missing context with tests.
Does Claude Code run locally?
In terminal and local IDE workflows, the client and approved tools operate on your machine. Model inference is still networked: prompts, selected context and outputs go to the active provider. Web sessions instead execute in Anthropic-managed cloud virtual machines.
Where can I use Claude Code?
Anthropic documents terminal, VS Code, JetBrains, Desktop and web surfaces. They use the Claude Code product layer, but authentication, execution location, settings behavior and feature availability are not identical.
Do I need a Claude subscription, and how much does Claude Code cost?
The answer depends on the surface and login. The terminal supports eligible Claude subscriptions, Anthropic Console usage and supported cloud providers; web, Desktop and managed services have separate plan rules. Prices, usage limits and seats change, so check Anthropic's live plan and provider pages instead of relying on a fixed monthly figure.
How do Claude Code permissions and sandboxing differ?
Permissions decide which tools and paths are allowed, denied or require approval. The optional sandbox adds OS-level filesystem and network restrictions for Bash commands and their child processes. Use both where the threat model requires defense in depth.
What do CLAUDE.md and MCP do in Claude Code?
`CLAUDE.md` supplies persistent project or organization instructions as model context; it is not a hard security policy. MCP connects external tools and data. Each server, credential and tool needs its own trust and permission review.
Does Anthropic train on or store Claude Code data?
It depends on the account, provider, privacy choice and surface. Anthropic documents opt-in controls for consumer model improvement, different commercial terms, local plaintext transcript caching and separate cloud-session data flows. Check the current settings and contract for the exact repository before use.
Who is responsible for reviewing Claude Code output?
The user and repository owner remain responsible. Claude can propose tests and review findings, but you should independently reproduce the behavior, inspect the diff and commands, run the normal quality and security gates, and retain human authority over merge and deployment.
Is Claude Code's GitHub Code Review available on every plan?
No. Anthropic's current feature matrix lists the managed GitHub Code Review service for Team and Enterprise. The local `/code-review` command is a separate CLI workflow, and neither service should be treated as approval to merge.