Claude Code

Claude Code is an agentic CLI that reads a codebase, edits files, runs shell commands, and drives git — installed via a native installer, Homebrew, or (deprecated) npm. The same engine now runs across the terminal, a VS Code/Cursor/JetBrains extension, a standalone desktop app, a browser (claude.ai/code), and CI via GitHub Actions or GitLab CI/CD, all reading the same CLAUDE.md, settings.json, and MCP server config, so a team's conventions travel with the repo rather than the tool. It also supports multi-agent orchestration (parallel subagents, an 'agent team' mode, and a scriptable Agent SDK for building custom agents on the same engine).

There are two independent paths, and mixing them up is the most common pricing mistake. (1) Subscription usage: Claude Code is included in every paid Claude plan and draws from the same rolling 5-hour/weekly usage pool as claude.ai chat — Pro is $17/month billed annually ($20 monthly), Max is $100/month (5x Pro's usage) or $200/month (20x), Team is $20/seat/month annually ($25 monthly, Standard) or $100/seat (Premium, 5x), and Enterprise is $20/seat plus usage at API rates. (2) Pay-per-token API: through a Console account, Claude Code bills per token at Sonnet 5's $2/MTok input and $10/MTok output (Opus 5: $5/$25; Haiku 4.5: $1/$5), separate from any subscription. Anthropic's own published enterprise figures put average cost at about $13/developer/active-day and $150-250/developer/month, with 90% of users under $30/day — a vendor claim, not independently verified.

The public changelog shows releases roughly every 1-3 days (2.1.259 through 2.1.267 across a two-week window in the sample reviewed), each with dozens of dated fixes across the CLI, VS Code extension, and gateway — not a marketing cadence but a genuinely fast-moving product. That cuts both ways: features you rely on can change quickly (e.g. npm install was deprecated in favor of a native installer during this period), and pinning a minimumVersion or release channel matters more here than for a slower-moving tool.

github.com/anthropics/claude-code has 144,571 stars and 23,078 forks, but it is not the product's source: its README states the CLI is installed via a native installer or npm, its package.json links back to a separate private repo (claude-cli-internal), and the npm package itself ships as a small wrapper plus platform-specific compiled binaries, not readable JS/TS source. What the public repo actually contains is the issue tracker, example plugins, MDM/devcontainer configs, and CI workflows. The license file is a one-line pointer to Anthropic's Commercial Terms of Service, not an OSS license.

Claude Code defaults to read-only in Manual mode and prompts before edits, command execution, or network calls; an 'auto mode' instead routes actions through a separate classifier model. It supports OS-level sandboxing for Bash (filesystem and network isolation) and a scoped working-directory boundary. Anthropic publishes SOC 2 Type 2 and ISO 27001 documentation through its Trust Center, and Claude Code has a HackerOne bug-bounty program. Anthropic explicitly does not security-audit third-party MCP servers a user connects — that risk sits with whoever configures them.

Public benchmarks like SWE-bench Verified and Aider's polyglot leaderboard score the underlying Claude models (Sonnet, Opus), not the Claude Code harness itself, so there is no independent, apples-to-apples benchmark of 'Claude Code' as a product against competing agentic CLIs (e.g. Aider, OpenAI Codex CLI, Cursor's agent mode) that we could locate. Anthropic's status page (status.claude.com) shows Claude Code sharing infrastructure incidents with claude.ai and the API — a September 3, 2026 elevated-error incident, for example, affected claude.ai, the API, Claude Code, and Cowork simultaneously — so outages are not isolated to one surface.