Vibeleaderboard
Index / tool
Visit code.claude.com
Category
Developer Tools
Rank
No. 01Tools index
Listed in
#1 Code with an agent
Pricing
Freemium
Platform
cli · desktop
Type
TOOL
Builder
anthropics
Latest release
v2.1.233
Added
Apr 14, 2026

About

Anthropic's agentic coding harness — the harness developers use to build and run autonomous coding agents from the terminal. It also works as an AI coding assistant that understands your entire codebase, executing routine tasks, explaining complex code, and handling git workflows through natural language.

What it does

This is the public repository for Anthropic's command-line coding agent, but the agent itself is not built here: the binary installs separately, and the repository carries everything around it. That includes thirteen bundled plugins, hook and settings examples, deployment templates for fleet-management tools, a development container with a restrictive firewall, and the automation that runs the public issue tracker. The extension model is simple and file-driven: hooks are shell commands fired at lifecycle events, and commands are plain markdown prompts, with no build step anywhere.

Why it's ranked here

The ranking rests on adoption surface and operational seriousness rather than openness. This repository carries the machinery of a project being actively run, not merely published: automated issue-tracker hygiene with sensible human overrides, and vulnerability reports routed into a bug bounty program. The plugin marketplace, the fleet-deployment templates, and the guarantee that managed settings outrank anything a user sets are exactly the layer an organization needs before putting an autonomous agent on engineers' machines. The proprietary license is the cost attached to all of it.

What's good

The enterprise controls are concrete rather than aspirational. A ready-made strict settings example disables the permission bypass mode, forces shell commands through an approval prompt, denies web access, and restricts hooks and permission rules to managed sources: an administrator can paste it and get a locked-down agent, then relax from there. The bundled development container drops all network traffic by default, allows only a short hard-coded list of hosts, and verifies itself by trying to reach an outside site and failing if that succeeds. The plugin format is legible: a directory, one manifest, and markdown.

Tradeoffs

The license is proprietary, all rights reserved, and the agent this repository configures arrives as an installed binary, so you cannot read or audit the thing you are adopting, only its configuration surface. The hook layer deliberately fails open: any error while evaluating a rule lets the tool call through instead of blocking it. Rule pattern matching is always case-insensitive with no way to opt out, and only a bare asterisk works as a wildcard when matching tools. The fleet-deployment templates carry their own warning that they are community-maintained and may be unsupported or incorrect.

How to use it well

Treat hooks as guidance and the permission system as the enforcement boundary: put anything you actually need blocked in the deny rules, because a rule engine that never fails closed will not hold. Start from the strict settings example and loosen it, rather than starting permissive. If you run a fleet, deploy settings through device management, where managed configuration sits at the top of the precedence order and users cannot override it. For unattended runs, use the bundled container and add your own package hosts to its firewall allowlist, since the default list covers only a handful of vendors.

Technical notes+

plugins/hookify/hooks/hooks.json registers one shell command per lifecycle event (PreToolUse, PostToolUse, Stop, UserPromptSubmit), each with a ten second timeout. plugins/hookify/hooks/pretooluse.py reads the tool call as JSON on stdin, loads rules from markdown files with YAML frontmatter, and prints a JSON decision on stdout; it catches every exception, emits the error as a systemMessage, and per its own comment always exits zero, so hook failures let the call through. plugins/hookify/core/rule_engine.py returns permissionDecision deny for blocking rules, compiles all patterns with re.IGNORECASE (no opt-out, cache of 128), and splits tool matchers on the pipe with exact membership, so only a bare asterisk acts as a wildcard. examples/hooks/bash_command_validator_example.py shows the lower-level contract: exit 2 blocks the call and stderr returns to the model. plugins/feature-dev/commands/feature-dev.md is a markdown prompt orchestrating parallel subagents across seven phases. examples/settings/settings-strict.json sets disableBypassPermissionsMode, allowManagedPermissionRulesOnly, allowManagedHooksOnly, and an empty strictKnownMarketplaces list. .devcontainer/init-firewall.sh sets iptables INPUT, OUTPUT and FORWARD policies to DROP, allowlists GitHub meta ranges, registry.npmjs.org, api.anthropic.com, sentry.io, statsig and VS Code endpoints, and verifies itself by curling example.com. scripts/sweep.ts runs the issue lifecycle: staleness thresholds, an upvote exemption, a skip for issues a human commented on after labelling, and a dry-run flag.

Observed

License
Proprietary: all rights reserved by Anthropic, subject to its commercial terms of service
Source availability
The repository gives installation instructions only, with no build steps for the CLI itself; its contents are plugins, examples, scripts and CI configuration
Runtime requirement
Node.js 18 or newer
Installation
Install script, Homebrew, PowerShell or WinGet; the npm package is deprecated
Bundled plugins
13 plugins declared in the built-in marketplace, including code review, feature development, hook tooling and security guidance
Hook events supported
PreToolUse, PostToolUse, Stop and UserPromptSubmit, each with a ten second timeout
Plugin dependencies
The hook plugin requires Python 3.7 or newer and states it has no external dependencies

Read from README.md, LICENSE.md, SECURITY.md, .claude-plugin/marketplace.json, plugins/README.md, plugins/hookify/README.md, plugins/hookify/core/rule_engine.py, plugins/hookify/hooks/pretooluse.py, plugins/hookify/hooks/hooks.json, plugins/feature-dev/commands/feature-dev.md, examples/hooks/bash_command_validator_example.py, examples/settings/settings-strict.json, .devcontainer/init-firewall.sh, scripts/sweep.ts, examples/mdm/README.md.

What it can do

  • Execute routine coding tasks through natural language commands

    Natural language description of coding taskExecuted code changes or automated task completion

  • Explain complex code sections

    Code snippet or file from codebasePlain language explanation of code functionality

  • Handle git workflows through natural language

    Natural language git command descriptionExecuted git operations (commits, branches, merges, etc.)

  • Analyze and understand entire codebase context

    Project directory or codebaseContextual understanding for code suggestions and explanations

  • Generate code from natural language prompts

    Natural language description of desired functionalityGenerated code implementation

  • Provide terminal-integrated coding assistance

    Commands and queries entered in terminalDirect coding help and code modifications in terminal environment

Intel on Claude Code

More in Intel

Tags

automationharnessdevelopercodingclaudeterminalagentAnthropiccliaigitAgents & Harnesses

Media

Claude Code

Comments (0)

No comments yet

Indexed by a proprietary survey. Corrections welcome.