Curiosity · AI Model
Claude Code
Claude Code is an agent product, not a raw model endpoint. Shipped by Anthropic in February 2025, it runs in the developer's terminal, reads and edits files directly, spawns subagents, and uses Model Context Protocol servers for external tools. Under the hood it routes to Claude Sonnet and Opus revisions, letting teams upgrade the brain without changing the CLI.
Model specs
- Vendor
- Anthropic
- Family
- Claude agents
- Released
- 2025-02
- Context window
- 1,000,000 tokens
- Modalities
- text, code
Strengths
- Deep integration with Anthropic's latest Claude models
- First-class MCP support and tool ecosystem
- Safe-by-default permissioning around file writes and shell commands
Limitations
- Not a standalone model — quality depends on the backing Claude revision
- Terminal-first UX can be unfamiliar to non-engineers
- Costs track the underlying Sonnet/Opus model usage
Use cases
- Repo-scale refactors driven from the terminal
- Agentic bug fixes with native git and shell tool use
- Scripted engineering workflows via MCP servers
- Classroom exercises on agent loops and tool use
Benchmarks
| Benchmark | Score | As of |
|---|---|---|
| SWE-bench Verified (Opus backend) | ≈75% | 2026-04 |
| Terminal-Bench | top-tier | 2025-12 |
Frequently asked questions
What is Claude Code?
Claude Code is Anthropic's official agent product that runs in the developer's terminal. It drives Claude models (typically Sonnet or Opus) through a tool loop that edits files, runs shell commands, calls git, and talks to Model Context Protocol servers.
Is Claude Code a model or a product?
Claude Code is a product — a CLI agent. Its intelligence comes from whichever Claude model it routes to (today, Claude Opus 4.7 by default). Upgrading the backing model upgrades the agent.
How does Claude Code handle security?
Claude Code enforces permission prompts for file writes, shell commands, and network calls by default. Teams can pre-grant specific permissions in settings, and sensitive actions like destructive git operations always require user confirmation.
Sources
- Anthropic — Claude Code overview — accessed 2026-04-20
- Anthropic — Claude Code announcement — accessed 2026-04-20