Claude Code

Tool reference

Built-in tools, permission rules, hook matchers, Bash/Edit/Grep/Agent behavior, and extending tools with MCP.

Claude Code tool names appear in permissions, hook matchers, subagent tools, skill allowed-tools, and CLI flags. Understanding the names is essential to reliable permission and automation rules.

#Common built-in tools

Tool	Purpose	Permission behavior
`Read`	Read files	Governed by Read deny/allow
`Edit`	Exact replacement in a file	Usually requires approval
`Write`	Create or overwrite files	Usually requires approval
`Bash`	Run shell commands	Usually requires approval
`PowerShell`	Run PowerShell on Windows or when enabled	Usually requires approval
`Glob`	Find filenames	Normally no prompt
`Grep`	Search contents with ripgrep	Normally no prompt, subject to Read path rules
`LSP`	Definitions, references, type diagnostics	Normally no prompt
`Agent`	Start a subagent	Launch normally has no prompt; internal tools retain checks
`Monitor`	Monitor commands, logs, or WebSockets in the background	Usually requires approval
`WebFetch`	Read web pages	Usually requires approval
`WebSearch`	Search the web	Usually requires approval
`Skill`	Invoke a skill	Depends on skill and rules
`Workflow`	Run a dynamic workflow	Usually requires approval

MCP tool names use mcp__server__tool. A mcp__* rule can target all MCP tools.

#Permission rule formats

Format	Tools covered	Example
`Bash(pattern)`	Bash, Monitor	`Bash(npm run *)`
`PowerShell(pattern)`	PowerShell	`PowerShell(Get-ChildItem *)`
`Read(path)`	Read, Grep, Glob, LSP	`Read(./docs/**)`
`Edit(path)`	Edit, Write, NotebookEdit	`Edit(./src/**)`
`Skill(name *)`	Skill	`Skill(deploy *)`
`Agent(name)`	Agent	`Agent(Explore)`
`WebFetch(domain:host)`	WebFetch	`WebFetch(domain:docs.example.com)`
`WebSearch`	WebSearch	`WebSearch`

An Edit allow rule usually also grants reads for the same path. Hook matchers use plain tool names, such as Edit|Write|Bash, not permission patterns such as Edit(./src/**).

#Bash behavior

Rule	Detail
Each call starts a separate process	Exported variables do not automatically persist to the next command
cd can persist in the main session	Within the project or additionally authorized directories
Subagents do not inherit Bash cwd	Each uses its own working directory
Default timeout is about two minutes	Claude can request longer; environment variables can configure it
Long output is truncated	Full output may be saved in session files for later reading
Long-running tasks can run in the background	Inspect or stop them with `/tasks`

For persistent variables, configure the shell before launching Claude, or use CLAUDE_ENV_FILE / a SessionStart hook.

#Edit behavior

Edit performs exact string replacement, not a fuzzy patch. It checks:

Check	Meaning
Read-before-edit	The session must have read the file, without an external change since
Match	old_string must match exactly
Uniqueness	Usually one match; otherwise provide more context or replace_all

Reading through Bash does not always count. Simple cat, sed -n, head, or tail may qualify, while pipes, redirects, and scripted reads usually do not. Prefer Read when in doubt.

#Grep, Glob, and LSP

Tool	Notes
Glob	Finds names, sorts by modification time, and may include gitignored files
Grep	Uses ripgrep, respects .gitignore by default, and follows rg regex syntax
LSP	Requires the language's code-intelligence plugin

In large repositories, locate symbols with Grep/LSP before reading a few key files. This reduces context usage and improves cache reuse.

#Agent tool

Agent launches subagents with separate context windows and caches. They return final results to the main session.

Scenario	Recommendation
Broad research	Isolate file reads in a subagent
Result needed immediately	Foreground subagent
Independent parallel work	Background subagent or /batch
Sensitive tools	Restrict tools or disallowedTools in subagent frontmatter

Launching a subagent normally does not prompt, but its Bash/Edit/Write operations still follow permissions. Use worktrees to isolate file changes.

#Extend tools with MCP

Built-in tools handle editing and local execution. Use MCP for external systems:

Need	Approach
Databases, Sentry, Linear, GitHub	Connect an MCP server
Many tools	Use supported tool search or split servers
Restrict an MCP tool	`deny: ["mcp__server__tool"]`
Observe calls	Match `mcp__server__tool` in a hook

See MCP integration.

#Official references

PreviousPlugin marketplaces and distribution NextStatus line

Support

Need help?

For setup, billing, or model issues, email us. Check the status page for uptime.

Support Email[email protected]Service Statusstatus.passion8.cc

WeChat / QQ support is available at the bottom right.