Claude Code

Terminal-first coding agent with strong repo reasoning, permission controls, project instructions, skills, plugins, and MCP workflows.

Type: CLI / IDE / web / desktop

Best for: Deep repository work, refactors, review, automation, and terminal-first engineering

Official docs: https://code.claude.com/docs/en/quickstart

Last verified: 2026-08-09

Status: Documentation verified

Verification scope: Official Claude Code quickstart, CLI reference, commands, permissions, tools, and MCP docs were reviewed. Commands were not executed locally in this repository.

60-Second Cheat Sheet

Need Command or file Notes
Install on macOS/Linux/WSL `curl -fsSL https://claude.ai/install.sh \ bash` Native installer; auto-updates
Install on Windows PowerShell `irm https://claude.ai/install.ps1 \ iex` Native Windows path
Start an interactive session claude Run from the repository root
Start with a task claude "explain this project" Opens an interactive session with the prompt
One-shot output claude -p "explain this function" Prints a response and exits
Start in plan mode claude --permission-mode plan Analyze and plan before implementation
Continue latest conversation claude -c Continues the latest conversation in the current directory
Resume a named/session thread claude -r "<session>" Resume by ID or name
Create repo instructions /init Generates a starter CLAUDE.md
Inspect/edit project memory /memory Manage CLAUDE.md and memory
Manage permissions /permissions Review allow, ask, and deny rules
Manage MCP connections /mcp or claude mcp External tools and context
Shared repo instructions CLAUDE.md Keep build, test, conventions, and safety guidance here
Shared MCP config .mcp.json Project-scoped MCP servers can be committed

Safe default: begin unfamiliar or high-impact work in plan mode, review proposed commands and edits, and widen permissions only for the specific task that needs them.

Five Prompts Worth Bookmarking

Understand a repository

Map this repository. Identify the main packages, entry points, build and test commands, and the safest first area to change. Do not edit files yet.

Plan a feature

Read the relevant implementation and tests, then propose a step-by-step implementation plan with files to change, validation commands, risks, and rollback points. Wait for approval before editing.

Review a diff

Review the current changes for correctness, regressions, security issues, and missing tests. List findings by severity and cite the affected files. Do not rewrite code unless I ask.

Fix a failing test

Reproduce the failing test, identify the smallest root-cause fix, explain why it failed, implement only that fix, and rerun the narrowest relevant validation first.

Refactor safely

Refactor this area without changing externally observable behavior. Preserve public APIs, add or update tests first where useful, and show the validation results before proposing follow-up cleanup.

What Claude Code Is Best At

  • terminal-first repository exploration and implementation
  • multi-file tasks that benefit from planning and iteration
  • repository onboarding and architecture explanation
  • review-heavy workflows with explicit permission controls
  • reusable project instructions through CLAUDE.md
  • extending local workflows with skills, plugins, hooks, subagents, and MCP

Installation

macOS, Linux, and WSL

curl -fsSL https://claude.ai/install.sh | bash

Windows PowerShell

irm https://claude.ai/install.ps1 | iex

Windows CMD

curl -fsSL https://claude.ai/install.cmd -o install.cmd && install.cmd && del install.cmd

The official quickstart also documents Homebrew, WinGet, and Linux package-manager paths. Native installations automatically update; package-manager installs should be updated through their package manager.

First Session

cd /path/to/your/repository
claude

A good first request is read-only:

What does this project do, where are the main entry points, and which commands build and test it? Do not edit anything.

Claude Code reads project files as needed, so you normally do not need to manually attach the repository.

Daily CLI Commands

claude
claude "fix the build error"
claude -p "explain this function"
claude -c
claude -r "auth-refactor" "finish the review"
claude --permission-mode plan

Useful in-session commands include:

/help
/init
/memory
/model
/permissions
/mcp
/agents
/clear
/resume

Use /help or / on your installed version for the authoritative current command list.

Repository Instructions

Use CLAUDE.md for durable repository guidance such as:

  • build, test, lint, and format commands
  • repository architecture and important directories
  • coding and review conventions
  • files or directories that should not be edited casually
  • validation expectations before a task is considered complete

Run /init for a starter file and /memory to inspect or refine memory and instruction files.

Keep CLAUDE.md concise enough that developers can review it like code. Context is guidance, not a security boundary.

Permission Model

Claude Code separates model guidance from enforcement.

Current official docs describe:

  • allow, ask, and deny permission rules
  • permission modes including default, acceptEdits, plan, auto, dontAsk, and bypassPermissions
  • OS-level sandboxing for Bash filesystem and network access
  • project-distributed permission settings

A practical team starting point is:

  1. use plan for unfamiliar, broad, or risky work
  2. keep sensitive files and dangerous commands denied
  3. allow only repeatable low-risk commands that your workflow genuinely needs
  4. use sandboxing as defense in depth rather than relying on prompt instructions alone

Avoid --dangerously-skip-permissions / bypassPermissions as a normal developer default.

MCP and External Tools

Claude Code supports project and user-scoped MCP servers.

A project-scoped server can be configured with claude mcp and stored in .mcp.json so the team shares the same server definition. Claude Code prompts before using project-scoped servers from a repository.

Treat every MCP server as a separate security decision because it can expand access beyond the local repository.

Review:

  • authentication method
  • data exposed to the server
  • available tools and write operations
  • network boundary
  • revocation path

Team Adoption Checklist

  • [ ] standardize where CLAUDE.md lives and who reviews it
  • [ ] document the default permission mode for normal work
  • [ ] define deny rules for secrets and destructive commands
  • [ ] approve MCP servers separately from Claude Code itself
  • [ ] require project tests or build validation before merge
  • [ ] keep bypass-permission workflows restricted to explicitly controlled automation

Security Notes

  • shell and filesystem access are the primary local risk boundary
  • permission rules and sandboxing are complementary controls
  • project instructions can improve behavior but do not enforce policy
  • MCP servers expand the tool and data boundary
  • keep credentials out of repository instructions and prompts

Alternatives

  • OpenAI Codex for OpenAI-centered terminal and multi-surface workflows
  • Cursor for IDE-first teams with a growing terminal agent surface
  • GitHub Copilot for GitHub-native editor, CLI, review, and repository workflows
  • Gemini CLI for a Google-backed terminal alternative with sandbox and workspace-context controls

See the Comparison Matrix for a side-by-side shortlist.

Verification Status

  • Status: Documentation verified
  • Last verified: 2026-08-09
  • Scope: install, CLI commands, project instructions, permissions, tools, and MCP reviewed against current official docs
  • Not locally tested: installer execution, account login, IDE/web/desktop behavior, and enterprise policy rollout

Sources