Claude Code

Anthropic’s coding agent: a terminal program, also available in IDEs, a desktop app and the cloud, that drives Claude models through a loop of reading, running and editing.

What it is: AI coding agents (terminal agents)
Type: Commercial product
Made by: Anthropic
License: Proprietary
Facts checked: 9 October 2026

What it is

Claude Code runs in your terminal, in a project directory. Anthropic describes its loop as three blended phases (gather context, take action, verify results) powered by a model that reasons and tools that act: reading and editing files, searching with patterns and regex, running shell commands and tests, fetching web pages, and spawning subagents. Each tool result goes back into the context and informs the next step. At startup it loads CLAUDE.md files from the directory tree (or a repository’s AGENTS.md in their place) plus its own auto memory, because every session otherwise starts with a fresh context window.

Two mechanisms keep it in check. Permission modes decide what runs without asking (Auto, Manual, Accept edits and a read-only Plan mode), and rules in .claude/settings.json allow, ask about or deny specific tool calls, checked in that order: deny, then ask, then allow. An optional sandbox, Seatbelt on macOS and bubblewrap on Linux, confines commands’ file writes and network access. Before each edit it snapshots the file, so you can rewind with Esc twice. When the context fills, it clears old tool output and then summarises the conversation; subagents do their searching in their own context and return only a summary.

The same loop runs in the terminal, VS Code and JetBrains extensions, a desktop app, the web, Slack and CI. Cloud sessions run on isolated, Anthropic-managed VMs and keep going after you close your laptop. Extensions plug in as MCP servers, skills, hooks and plugins, and the Agent SDK exposes the same tools for building your own agents.

Why teams pick it

A layered safety model: deny-first permission rules, an OS-level sandbox and file checkpoints you can rewind
One loop across terminal, IDE, desktop, web, Slack and CI, so a session can move with --cloud and --teleport
Composable on the command line: claude -p takes piped input, prints a result and exits, for scripts and CI
A deep extension surface (MCP, skills, hooks, subagents, plugins); MCP tool definitions load on demand to save context

Watch out for

Closed source, and it drives only Claude models, through Anthropic or Amazon Bedrock, Google Cloud and Microsoft Foundry
Most surfaces need a Claude subscription or an Anthropic Console account; cloud sessions count against plan usage limits
Rules match command text: a deny on curl doesn’t catch the same program by full path or inside sh -c, so only the sandbox makes a restriction hold
Compaction can drop detailed instructions from early in a session, and checkpoints can’t undo effects on remote systems such as databases or deployments

Alternatives

What all terminal agents have in common →

Gemini CLICommercialGoogle’s open-source terminal agent for Gemini models; since 18 June 2026 it serves only paid API keys and business licences, as Google moved free and consumer users to Antigravity CLI.

OpenAI CodexCommercialOpenAI’s coding agent: an open-source CLI, written in Rust, that works in your local repository inside an OS sandbox, plus an IDE extension and Codex cloud for tasks that run on OpenAI’s machines.

OpenCodeOpen sourceAn open-source (MIT) coding agent that works with 75+ model providers and local models, built as a local server with a terminal interface as one of its clients.+1 Kilo Code

How it fits with other technologies

What Claude Code uses

What uses Claude Code

Augment Codeserves its Context Engine over MCP toClaude CodeAI coding agents

Start a session, then script it · bash

cd my-repo && claude      # interactive; /init writes a starter CLAUDE.md

# one prompt, no session: prints the answer and exits
claude -p "explain what src/auth does"

# pipe input in, as in CI
git diff main --name-only | claude -p "review these changed files for security issues"

# read-only exploration before any edits
claude --permission-mode plan

Permission rules in .claude/settings.json#

With this configuration Claude Code runs npm scripts and git commits without asking and refuses commands that begin with git push. A push written another way, such as git -C . push, isn’t matched, which is why the docs pair rules with the sandbox.

json

{
  "permissions": {
    "allow": [
      "Bash(npm run *)",
      "Bash(git commit *)"
    ],
    "deny": [
      "Bash(git push *)"
    ]
  }
}

See it in a real stack

No architecture uses Claude Code yet. Drop it into the playground and wire it up yourself.Open playground →

Worth exploring

Video· August 2025Advanced Context Engineering for AgentsDex Horthy (HumanLayer), YC Root AccessA talk on getting coding agents to work in large, existing codebases by keeping their context small and correct: split work into research, plan and implement, compact each step into a Markdown file, keep context use around 40–60%, and spend human review on the research and plan rather than every line of code. Summarised here from the written version in HumanLayer’s GitHub repository, which also notes where the approach failed; HumanLayer builds tools for this workflow.

All resources →

Where it came from2024–2025

Anthropic introduced Claude Code on 24 February 2025, alongside Claude 3.7 Sonnet, as “our first agentic coding tool” and a limited research preview. It could search and read code, edit files, write and run tests, commit and push to GitHub and use command-line tools, keeping the developer in the loop at every step. Anthropic said its goal was partly to learn how developers use Claude for coding, to inform future models, and that its own team had found it most useful for test-driven development, debugging and large refactors.

Three months later, with Claude 4 on 22 May 2025, Claude Code became generally available. The same release added beta extensions for VS Code and JetBrains, background tasks through GitHub Actions, and the Claude Code SDK so others could build agents on the same core. Since then it has spread to a desktop app, the web, Slack and cloud sessions, while keeping one agentic loop across all of them.

Why it’s built this way

Start in the terminal: A terminal tool can use anything you can run from the command line, from build tools to git, and composes with pipes and CI. Anthropic describes it as helping with anything you can do from the command line, not only coding.
A thin harness around the model: Claude Code supplies tools and manages context; the model decides which tool to call next based on what the last one returned. That keeps the product close to the model Anthropic wanted to learn about.
Persistent instructions as Markdown files: Sessions start with an empty context, so CLAUDE.md gives a checked-in place for conventions that every session loads; later versions read AGENTS.md too, for repositories shared with other agents.
Ask by default, then loosen: An agent that runs commands needs a way to stop it. Permission modes, rules and checkpoints let a team start with approvals and grant autonomy command by command.

Timeline

2024
Anthropic open-sources the Model Context Protocol, later Claude Code’s way to add tools (November)
2025
Research preview launches with Claude 3.7 Sonnet (February)
2025
Generally available with Claude 4; IDE extensions, GitHub Actions and the SDK (May)
2025
Anthropic donates MCP to the Linux Foundation’s Agentic AI Foundation (December)

Sources11 · facts checked 9 October 2026

Improve this page on GitHub