Claude Code Tools

Read loads file contents so Claude Code can understand your existing code before making changes. It handles text files, images, PDFs, and Jupyter notebooks, and supports reading specific line ranges for large files. Claude Code reads files automatically whenever it needs context — you rarely need to ask for it explicitly.

Write creates new files from scratch or fully replaces the contents of an existing one. It's used when Claude Code needs to generate something entirely new — a config file, a new component, a migration script. For modifying existing files, Edit is preferred since it targets specific sections without touching the rest.

Edit performs exact string replacements in files — find the old text, replace it with new text. This is how Claude Code makes surgical changes to your code without rewriting entire files. It preserves everything else in the file exactly as-is, which means your formatting, comments, and surrounding code stay untouched.

Glob searches for files using patterns like **/*.tsx or src/**/test.*.js. It's fast even on huge codebases and returns results sorted by modification time. When Claude Code needs to find where something lives — a component, a config file, test files for a module — Glob is usually the first step.

Grep searches inside files for text matching a pattern — function names, error messages, imports, anything. It supports full regex, file type filtering, and context lines around matches. When you need to trace how something is used across a codebase, Grep finds every reference in seconds.

NotebookEdit modifies individual cells in .ipynb files — replace a cell's contents, insert a new cell, or delete one. It works with both code and markdown cells. This is how Claude Code helps with data science workflows without mangling the notebook's JSON structure.

Bash executes shell commands in your terminal, giving Claude Code the ability to run tests, install packages, manage git, spin up services, and do anything else you'd do from the command line. The working directory persists between calls, so multi-step workflows work naturally. Commands time out after 2 minutes by default, with an option to extend up to 10 minutes for longer operations.

WebFetch grabs the content of a web page, converts it to clean markdown, and processes it with a prompt to extract what you need. It's useful for pulling in documentation, reading blog posts, checking API references, or grabbing any public web content. Results are cached for 15 minutes, so repeated fetches are fast.

WebSearch runs a web search and returns results that Claude Code can use to answer questions about recent events, new library versions, or anything beyond its training data. It supports domain filtering to focus results on specific sites or exclude others.

Task creates a sub-agent that runs independently with its own context window, perfect for parallelizing work that doesn't depend on each other. Need to research three different libraries, refactor two unrelated modules, or explore a large codebase? Spin up sub-agents and let them work simultaneously instead of doing everything sequentially.

TaskCreate builds a task list for complex, multi-step projects. Each task gets a subject, description, and active status label. This is how Claude Code organizes large refactors, multi-file changes, or any work that benefits from a visible progress tracker.

TaskGet fetches the complete details of a task by its ID — subject, description, status, and dependency information. It's used to understand exactly what a task requires before starting work on it, especially in team settings where tasks are assigned by a lead.

TaskUpdate modifies task properties — mark it in-progress, completed, or deleted, change the description, add or remove dependencies. It's the primary way Claude Code tracks progress through a task list and signals to you (or teammate agents) what's done and what's next.

TaskList shows every task in the current session — pending, in-progress, and completed — along with their dependencies and assigned owners. It's how Claude Code decides what to work on next and how it shows you overall progress on a multi-step project.

TaskOutput retrieves results from tasks running in the background — sub-agent research, long-running shell commands, or any async operation. It lets Claude Code check on work that's still running or collect results from work that's finished.

TaskStop terminates a background task that's no longer needed. If a sub-agent is doing research you no longer care about, or a background build is running against the wrong branch, TaskStop kills it so resources aren't wasted.

TeamCreate assembles a group of Claude Code agents, each with a defined role and set of capabilities. One agent might handle frontend components while another works on the API layer and a third writes tests. They coordinate through the task system and messaging, working simultaneously instead of sequentially.

TeamDelete shuts down a multi-agent team when the work is done. It sends shutdown signals to all agents on the team and cleans up resources. This is typically called by the team lead once all tasks are completed and results have been collected.

SendMessage is how agents on a team communicate — sharing progress updates, asking questions, reporting blockers, or coordinating handoffs. It supports direct messages to specific teammates and broadcasts to the whole team. This keeps multi-agent workflows coordinated without agents overwriting each other's work.

EnterPlanMode puts Claude Code into a planning state where it outlines an approach before writing any code. This is valuable for complex changes where you want to review the strategy first — architectural decisions, multi-file refactors, or features that touch several systems. No code gets written until you approve the plan.

ExitPlanMode presents the completed plan for review. In a solo session, the user approves or requests changes. In a team setting, the team lead reviews the plan. Once approved, Claude Code switches from planning to implementation and starts writing code according to the agreed-upon approach.

AskUserQuestion pauses work to ask you something when Claude Code doesn't have enough information to make a confident decision. Instead of guessing and potentially going down the wrong path, it asks — with suggested options when applicable. This saves time by catching misunderstandings early.

Skill triggers a pre-defined workflow that's been registered with Claude Code. Skills like /commit, /review-pr, or custom project-specific skills package up common multi-step operations into a single command. They're shortcuts for things you do repeatedly.