Anthropic's agentic terminal — a coding agent that reads entire codebases, plans multi-file changes, edits, runs tests, and iterates on failures. The same engine the Claude Agent SDK runs on, with a tightly-curated CLI surface. Bundled with Claude Pro, Max, Team, and Enterprise subscriptions; locked to Claude (Sonnet / Opus) as the model.
Not autocomplete. Not chat-with-code. A persistent agent that holds the whole codebase in mind, plans changes across files, executes them, runs your tests, reads the output, and iterates until things pass.
Claude Code launched as Anthropic's answer to the "AI pair programmer" framing — but pitched a level above. The agent doesn't suggest the next line; it commits to a multi-step plan ("read these six files, refactor the auth middleware, update the three call sites, run the tests, fix the failures") and works the plan to completion. Approval-by-default at every destructive step keeps a human in the loop.
It runs in the terminal first — that is the primary product surface — with VS Code, JetBrains, and a claude.ai/code web app available for teams that prefer those entry points. Same engine across all surfaces; only the affordances differ.
Two production receipts that come up in every comparison: Stripe migrated 10,000 lines of code in 4 days using Claude Code, against a 10-week manual estimate. Wiz refactored a 50,000-line library in roughly 20 hours. These are real, vendor-cited outcomes; the savings are not theoretical.
Six load-bearing capabilities. Each is shipped, documented, and used in production.
Reads the project, traces dependencies, performs multi-file refactors. Not "this file"; the whole graph.
Runs tests, reads error output, fixes code, re-runs. Stops when green or the budget hits. Plan mode requires step-by-step approval; Autopilot mode lets it run.
Approval-by-default before file mutations or shell commands. Configurable per-project. The agent will not write a destructive command without confirming.
Skills are markdown instructions in ~/.claude/skills/. Hooks fire on git events. MCP integrates 3,000+ external tools (databases, GitHub, Sentry, etc.). All three compose.
Delegate tasks to specialised agents with isolated context, custom prompts, and restricted tool access. Prevents context bloat on long tasks.
Plan mode guides you through task analysis before execution. Autopilot mode iterates autonomously through the plan. Switch per task with a slash command.
All four call the same Anthropic API and run the same agent loop. Pick the surface that matches how the team works.
The canonical experience. claude launches the agent in your shell with full repo access. Most features land here first.
Side-panel chat + inline diffs. Skills and MCP servers carry over from the terminal install. Best for teams already in VS Code.
IntelliJ, GoLand, PyCharm, etc. Feature parity is close but slightly behind the terminal.
Browser-based sandbox for quick experiments without local setup. Limited compared to terminal but useful for exploration.
Claude Code is not a standalone product — it's included with paid Claude tiers. The interesting price signal is the upcoming separation of programmatic SDK usage from chat credits.
Claude Pro — $20/month (or $17/month annual). Includes Claude Code at standard usage limits. The default starting point.
Claude Max 5x — $100/month. Claude Code with 5× usage multiplier. For heavy individual users.
Claude Max 20x — $200/month. 20× usage. For developers running Claude Code most of the day.
Team and Enterprise — per-seat with admin controls, SSO, audit log retention. Pricing on request.
The free Claude tier (web + mobile chat) does not include Claude Code.
From 2026-06-15, programmatic SDK usage (i.e. Claude Code calls beyond the chat experience) starts drawing on a separate credit pool:
— Pro: $20/month SDK credits
— Max 5×: $100/month SDK credits
— Max 20×: $200/month SDK credits
This separation effectively raises the cost ceiling for heavy programmatic use. Budget it before the date if your team uses Claude Code intensively.
Honest tradeoffs. Single-model is its strength and its limit.
Pricing. At ZAR/USD rates, Pro at $20/month is roughly R380/month per developer (mid-2026). For a 10-person team that's ~R45k/year. Max 20× at $200/month is R3,800/month per seat — only worth it for developers running Claude Code most of the day.
Local model alternative. Claude Code does not support local inference. For "code never leaves the building" environments (regulated industries, sensitive client work), pair it with the API tier and a region that meets your residency constraint, or use an OSS alternative like Aider against Ollama instead.
Talent. Most SA engineers in 2026 have used Cursor or Copilot before Claude Code. Plan a sprint or two of pair-programming for onboarding — the model shift (Claude vs whatever they're used to) and the CLI-first surface (vs IDE-first Cursor) both take some adjustment.
POPIA. Anthropic offers data-residency commitments at the Enterprise tier; consult the current terms before deploying on customer data. Free / Pro / Max tiers do not carry the same guarantees.
Claude Code moves fast — verify versions against the date on this leaf.