Claude Code is an agentic coding tool from Anthropic. It reads your codebase, edits files, runs commands, and executes multi-step development tasks guided by natural language. It is available as a terminal CLI, a VS Code extension, a JetBrains plugin, a desktop app, and via the web at claude.ai.
The core interaction model is conversational: describe a task, and Claude Code reads the relevant parts of the codebase, plans the work, and executes it. This differs from autocomplete-based tools, which operate on the current file and suggest completions as you type. Claude Code operates at the project level — it can read across many files, run shell commands, manage git, and make coordinated changes without manual file-by-file direction.
Quick Facts
Version
2.1.139 (May 11, 2026)
Pricing
Pro $20/month · Max 5x $100/month · Max 20x $200/month
Team pricing
Team Premium $100/seat/month (min. 5 seats)
SDLC stage
Write / Build
Platforms
Terminal, VS Code, JetBrains, Desktop app, Web
Models
Sonnet 4.6 (default), Opus 4.7 (Max+)
Context window
1M tokens in Claude Code (Opus 4.7 and Sonnet 4.6; automatic on Max, Team, Enterprise; Pro requires extra usage enabled)
MCP support
Yes
API access
Yes — Sonnet 4.6 at $3/$15 per million tokens
What It Does
Claude Code handles software development tasks that span multiple files and require understanding how pieces of a codebase relate to each other. Its primary applications include:
Full-codebase refactoring — renaming interfaces, migrating SDKs, restructuring modules across many files simultaneously
Automated test generation — reading existing implementations and writing tests against them
Code review and explanation — tracing request paths, explaining unfamiliar codebases, surfacing why decisions were made
Bug investigation — running grep, git log, and file reads as part of a single diagnostic session
Claude Code executes these tasks through an agent loop: it reads files, runs commands, makes edits, and checks results — repeating until the task is complete or it needs input. Permission modes control how much it can do autonomously. In the default mode (default), it prompts for approval before each file edit or shell command. In plan mode (plan), it presents the full execution plan before starting, rendered as a markdown document where you can add inline comments before execution begins. In acceptEdits mode, it applies file edits without pausing but still prompts for shell commands.
For infrastructure-as-code work, Claude Code can generate complete module configurations including variables, outputs, and provider-specific arguments. Example of a Terraform Aurora PostgreSQL cluster:
CLAUDE.md is the primary configuration surface. It is a project-level file where teams document conventions, directory structure, architectural decisions, and commands. Claude Code reads it at session start and caches it across the session. Example:
# CLAUDE.md — infrastructure monorepo
## Repository structure
- /modules — reusable Terraform modules
- /services — individual service deployments
- /legacy — read-only, do not modify
## Conventions
- Terraform follows Google style guide
- State backend: S3 + DynamoDB per environment
- PR titles follow conventional commits
## Commands
- `make plan ENV=staging` — Terraform plan
- `make test` — run Terratest suite
Slash commands available in all interfaces:
/init # generate CLAUDE.md from repo structure
/compact # summarize context to free window space
/model # switch between Sonnet and Opus
/cost # show current session token consumption
/clear # start fresh context
/goal # set a completion condition; Claude works until met
According to the v2.1.139 release notes, the /goal command allows setting a completion condition that Claude Code works toward across multiple turns, showing elapsed time, turn count, and token consumption as a live overlay.
Multi-agent workflows are supported via the Agent SDK. Claude Code instances can spawn sub-agents to work on parallel tasks, with each sub-agent carrying its own context and tool access. The agent view (claude agents) shows all running, blocked, and completed sessions in a single list.
Platforms and Integration
Terminal CLI — Installs via curl, Homebrew, or WinGet. Supports macOS, Linux, WSL, and native Windows. Auto-updates on native installs. The CLI includes the full command set, tab completion, bash shortcut (!), and access to all MCP configuration options.
VS Code extension — Provides a graphical panel with inline diffs, @-mentions with file references, plan review before accepting changes, conversation history, and multiple sessions in tabs. The extension works in Cursor, Windsurf, Kiro, and other VS Code forks. Installation: search "Claude Code" in the Extensions view or install from the VS Code Marketplace.
The VS Code extension supports checkpoints, which track Claude's file edits and allow rewinding to any prior state. Three rewind options are available per message: fork the conversation from that point while keeping code changes, revert file changes to that point while keeping conversation history, or do both simultaneously.
The Chrome extension (@browser command) connects Claude Code to your active Chrome tab, enabling it to read page content, check console output, and automate browser interactions as part of a coding session. Requires Claude in Chrome extension v1.0.36+ and Claude Code v2.0.73+.
JetBrains plugin — Available for IntelliJ IDEA, PyCharm, WebStorm, GoLand, Rider, CLion, and other JetBrains IDEs. Provides the same conversational interface as the VS Code extension within the JetBrains environment.
Desktop app — Standalone app for macOS and Windows, installable without a browser.
Web — Available at claude.ai for browser-based access without local installation.
Third-party model providers — The terminal CLI and VS Code extension support Amazon Bedrock, Google Vertex AI, and Microsoft Foundry as model providers in place of Anthropic's API directly. Configuration is in ~/.claude/settings.json.
MCP (Model Context Protocol) — Claude Code supports MCP servers for connecting external tools, databases, and APIs. Configure via claude mcp add in the terminal; manage existing servers with /mcp in the VS Code panel. Remote MCP servers over HTTP/SSE are supported in addition to local stdio servers.
CI/CD and code review — Claude Code runs in headless mode (claude -p) for integration into automated pipelines. Common applications include automated pull request review, test generation on commit, and documentation updates. The Agent SDK supports building multi-agent systems where Claude Code instances coordinate on parallel workstreams.
In Practice
The Pragmatic Engineer's February 2026 survey of 900+ software engineers ranks Claude Code as the most-used AI coding tool eight months after its May 2025 release, ahead of GitHub Copilot and Cursor. 46% of respondents name it the most-loved tool — more than double Cursor at 19% and over four times GitHub Copilot at 9%. Adoption is highest at smaller companies (75% of respondents at 2–50-employee companies vs. 49% at 10,000+). Among the 55% of respondents who report regularly using AI agents, 71% of that cohort name Claude Code as the tool of choice, and the most-cited agent use cases are code review and validation, bug fixing, code investigation, and automating repetitive tasks.
Repository activity is consistent with that adoption signal. The anthropics/claude-code GitHub repository carries 124,000 stars, 20,400 forks, 542 open pull requests, and over 5,000 open issues as of May 2026, with roughly 52 listed contributors and ~610 commits since the v1.0.0 GA release in May 2025 (per an Augment Code analysis published May 5, 2026). Release cadence is rapid: 11 tagged versions shipped between May 6 and May 13, 2026 — v2.1.131 through v2.1.141 — each carrying bug fixes and incremental features visible in the public release notes. MDM deployment templates were added to the public examples directory in April 2026.
A workshop write-up by Vincent Quigley, Staff Software Engineer at Sanity (sanity.io blog, September 2, 2025), describes a three-attempt iteration pattern in production use: a first pass with roughly 95% unusable output, a second pass at roughly 50%, and a third pass that produces workable code for review. The same post reports per-engineer Claude Code spend of $1,000–$1,500 per month for senior engineers using it heavily, and describes running multiple parallel Claude Code instances coordinated through Linear tickets, with project-level CLAUDE.md files connected via MCP to Linear, Notion, GitHub, and read-only databases.
On Hacker News, the "My experience with Claude Code after two weeks of adventures" thread (July 2025, 600+ comments) and the "Claude Code is all you need" thread (September 2025) repeatedly describe Claude Code displacing GitHub Copilot and being run inside Cursor or VS Code as a terminal pane rather than as a replacement for the editor itself. Recurring practitioner observations in those threads include uninterrupted agent runs of 10+ minutes without getting stuck in loops, and dynamic file discovery via grep and find rather than embedding-based codebase indexing as the differentiating behaviors versus IDE-embedded agents.
Pricing
Claude Code is included in Pro, all Max plans, and Team plans (Team Standard as of late April 2026). It is not included in the Free plan.
Individual plans:
Plan
Monthly
Annual
Claude Code
Free
$0
—
No
Pro
$20/month
$17/month ($200/year)
Yes
Max 5x
$100/month
—
Yes + Opus 4.7 + priority access
Max 20x
$200/month
—
Yes + higher usage limits
Team plans:
Plan
Annual
Monthly
Claude Code
Team Standard
$20/seat/month
$25/seat/month
Yes (added late April 2026)
Team Premium
$100/seat/month (5-seat min.)
$125/seat/month
Yes + SSO, SCIM, usage analytics
Team Standard includes Claude Code as of late April 2026. Previously it was Premium-only. Team plans support up to 150 seats; above that, Enterprise is required.
Enterprise — Custom pricing via Anthropic sales. Adds a 500K context window for general Claude use (vs 200K on Team), a 1M token context window in Claude Code, HIPAA readiness, custom data retention, and compliance tooling.
Usage operates on a rolling 5-hour token window rather than a monthly quota. The window resets every 5 hours. Intensive agentic sessions — large refactors or test generation runs with many tool calls — consume the window faster than conversational use. Max 5x provides approximately 5x the message capacity of Pro per window; Max 20x provides approximately 20x.
An extra usage toggle is available on all paid plans. When the window limit is reached, Claude Code can continue at standard API rates with a spend cap you set.
If ANTHROPIC_API_KEY is set as an environment variable, Claude Code bills the API account instead of the subscription. Teams using both should verify which credential is active before deploying to shared environments.
Extended thinking is on by default for complex tasks. Thinking tokens are billed at output token rates. Set MAX_THINKING_TOKENS=8000 to cap the number of thinking tokens per request.
API pricing (pay-per-token, no subscription required):
Model
Input
Output
Sonnet 4.6
$3/MTok
$15/MTok
Opus 4.7
$5/MTok
$25/MTok
Haiku 4.5
$1/MTok
$5/MTok
Prompt caching reduces repeated context costs. Cached reads are billed at 10% of standard input pricing. CLAUDE.md files and system prompts are automatically cached.
Requirements
Claude Pro, Max, or Team subscription — or API credits
# Native install (macOS/Linux/WSL) — auto-updates
curl -fsSL https://claude.ai/install.sh | bash
# Or via Homebrew (manual updates)
brew install --cask claude-code
# Verify version
claude --version
# Start in a project
cd your-project
claude
CLAUDE.md setup — Run /init in a new project to generate a starter CLAUDE.md from the repo structure. Edit it to add team conventions, out-of-scope directories, and build commands.
Security and Data
According to the official Claude Code security documentation, the following applies across plans.
Data training policy — For commercial users (Team, Enterprise, API, and third-party platform users), Anthropic does not train generative models on code or prompts sent to Claude Code unless the organization has explicitly opted in via the Development Partner Program. For consumer users (Free, Pro, Max), model training is opt-in and controlled by an account-level setting.
What runs locally vs. what is transmitted — File reads, shell command execution, git operations, and code edits all happen locally on the machine. The content of files read and the output of commands run are transmitted to the Anthropic API as part of the conversation context for the model to process. Known API key and token patterns are redacted before any transcript is uploaded.
Session feedback and transcripts — Transcripts are not uploaded automatically. If you respond "Yes" to the optional session quality survey follow-up ("Can Anthropic look at your session transcript?"), the conversation transcript and session log are uploaded and retained for up to 6 months. Transcripts submitted via the /feedback command are retained for 5 years. Organizations with zero data retention configured never see the upload prompt.
Permission model — In the default mode, Claude Code prompts for approval before each file edit or shell command. Six permission modes are available: default (prompts per action), acceptEdits (auto-accepts file edits, still prompts for commands), plan (read-only analysis, no file changes), auto (AI classifier replaces manual prompts; requires Team, Enterprise, or API plan), dontAsk (pre-approved tools only), and bypassPermissions (skips all checks, for isolated environments only).
Compliance — SOC 2 Type 2 report and ISO 27001 certificate are available at trust.anthropic.com. HIPAA readiness is available on the Enterprise tier.
Zero data retention — Available on Claude for Enterprise only. When enabled, prompts and responses are processed in real time and not stored by Anthropic after the response is returned. Details at code.claude.com/docs/en/zero-data-retention.
Alternatives
GitHub Copilot ($10/month Pro; transitioning to usage-based billing June 1, 2026) — Inline autocomplete as the primary feature, with agent mode for multi-file tasks. Supports JetBrains, Xcode, Neovim, and all major IDEs. According to the Pragmatic Engineer survey, Copilot sees its highest usage at large enterprises (10,000+ employees), which respondents attributed to Microsoft's procurement relationships. The June 2026 credit model means heavy agent mode users will see costs above the $10 headline price. Copilot's enterprise PR review agent integrates natively with GitHub Actions for automated pull request feedback within CI/CD pipelines.
Cursor ($20/month Pro, $60/month Pro+, $200/month Ultra) — An IDE built on VS Code with agentic coding features, visual diff review, and browser preview integration. Used by 35% of The Pragmatic Engineer survey respondents (Feb 2026), growing ~35% in nine months. The survey notes that teams that have been Cursor power users are introducing Claude Code while keeping Cursor in place, using both tools rather than switching. Cursor indexes the codebase using embeddings for fast context retrieval; Claude Code builds context dynamically at runtime via tool calls (grep, find, file reads), which affects how each tool handles large or frequently changing codebases.
OpenAI Codex (included in ChatGPT Pro) — OpenAI's agentic coding tool. According to the Pragmatic Engineer survey, Codex reached 60% of Cursor's usage in the months since launch, with particularly strong growth at smaller companies.
Gemini CLI (free, open source) — Google's terminal-based coding agent powered by Gemini models. Installable via npx. Integrates with Google Cloud services — including Cloud Run, BigQuery, and GCP IAM — directly from the command line. Supports embedding in CI/CD scripts and can be run in headless mode for pipeline automation, comparable to Claude Code's claude -p headless mode. Reported by ~11% of Pragmatic Engineer survey respondents, with usage stable across company sizes.
OpenCode (open source, MIT license) — Model-agnostic terminal-based coding agent. Accepts any model provider via configuration, allowing teams to route different task types to different models — for example, complex architecture queries to a larger model and routine tasks to a faster, lower-cost model. According to the Pragmatic Engineer survey, OpenCode is gaining adoption among Staff+ engineers. The model-agnostic architecture avoids vendor lock-in, which the survey respondents in that cohort cited as a factor in tool selection.
Claude Code
Cursor Pro
GitHub Copilot
Gemini CLI
Price
From $20/month
$20/month
From $10/month*
Free
Interfaces
Terminal, VS Code, JetBrains, Desktop, Web
IDE (VS Code-based)
All major IDEs
Terminal
Inline autocomplete
No
Yes
Yes (primary feature)
No
Multi-file agentic work
Yes
Yes
Partial
Yes
Visual diff / preview
Via VS Code extension
Built-in
Limited
No
JetBrains native plugin
Yes
No
Yes
No
MCP support
Yes
Yes
Partial
No
Open source
No
No
No
Yes
Model flexibility
Yes (via AWS/GCP/Azure)
Multi-model
GitHub models
Google only
SOC 2 Type 2
Yes
Yes
Yes
N/A
*Copilot moving to credit-based billing June 1, 2026.
Version & Freshness Metadata
Last reviewed: May 2026 · Version: 2.1.139 · Next review: August 2026
