Tools

Tools are how the agent interacts with your system. Each tool has a name, a sensitivity level, and a set of parameters. The model decides which tool to call based on the tool’s description and the current task.

How tool calls work

1. The model emits a tool call with a name and input arguments
2. mew checks permissions (see Permissions)
3. If approved, the tool runs and produces output
4. The output goes back to the model as context for the next turn

In the TUI, tool calls appear as cards inline in the conversation. Each card shows the tool name, its input, and the output. Bash output can be expanded or collapsed with Ctrl+O. Diffs (from Write and Edit) get syntax-colored +/- highlighting.

Built-in tools

File operations

Tool	Sensitivity	Description
read	ReadOnly	Read file contents. Returns a [path#hash] header and numbered lines for hashline edits. Supports offset and limit for pagination. Detects binary files and refuses to read them.
write	Mutating	Write content to a file. Creates the file if it doesn’t exist, overwrites if it does.
edit	Mutating	Replace a unique string in a file. Fails if the old string appears more than once. Includes first/last line snippets in error messages to help recovery.
edit_hashline	Mutating	Edit files using line-anchored hashline patches with file-hash staleness detection. Supports block-aware ops and stale-tag recovery.
glob	ReadOnly	Find files matching a glob pattern. Sorted by modification time. Honors .gitignore by default.
grep	ReadOnly	Search file contents for a regex pattern. Prefers ripgrep if available. Supports include filters and context lines.

See Hashline Edits for the patch format and examples.

Shell

Tool	Sensitivity	Description
bash	Dangerous	Execute a shell command. Output is captured and returned. Secret redaction applies to output.
shell_background	Dangerous	Launch a shell command in the background and return a job ID immediately.
shell_monitor	Dangerous	Run a command and wait for it to exit successfully, retrying until a timeout. For readiness probes.
job_status	ReadOnly	Check the status of a background job (shell or subagent).
job_block	Mutating	Wait for a background job to reach a terminal state.
job_cancel	Dangerous	Cancel a running background job by killing its process.

Session and agent

Tool	Sensitivity	Description
echo	ReadOnly	Echoes back the provided input. Used for testing and diagnostics.
exit_tool	ReadOnly	Stop the current subagent run and return a final answer to the parent. Top-level use ends the session.
flag_important	ReadOnly	Mark a file as important so it survives context compaction. Flagged files are kept in context even when older messages are summarized.
progress_update	ReadOnly	Report a status update to the parent agent. Used by subagents mid-run to show what they’re doing in the sidebar.
ask_user_question	ReadOnly	Ask the user 1-4 multiple-choice questions when their answer would change how to proceed. Renders an interactive question card.

Subagents

Tool	Sensitivity	Description
subagent_start	ReadOnly	Spawn a child agent with a fresh conversation context, tool allowlist, and system prompt. Returns a job ID.
subagent_wait	ReadOnly	Wait for a background subagent to complete and get its result.

See Subagents for how subagents work and how to define custom ones.

Todos

Tool	Sensitivity	Description
todo_create	ReadOnly	Create one or more todos. Each todo tracks a step of the work.
todo_update	ReadOnly	Update a todo’s content and/or status (pending, in_progress, done, blocked).
todo_complete	ReadOnly	Mark a todo done. Refused if dependencies aren’t done.
todo_delete	ReadOnly	Delete a todo. Refused if another todo depends on it.
todo_list	ReadOnly	List all todos with their IDs, statuses, and dependencies.

The todo list renders in a panel below the chat. Press Ctrl+P and select “Todo” to toggle its visibility, or use /todo to print it as text.

Web

Tool	Sensitivity	Description
web_fetch	ReadOnly	Fetch a URL and return its content as markdown. HTML pages are converted to readable markdown. Content is truncated at 128KB.

web_fetch is useful for reading documentation pages, API references, and articles without leaving the conversation. It follows up to 5 redirects and sets a 30-second timeout.

Conditional tools

These tools only register when their prerequisites are met:

Tool	Condition	Description
skill	At least one skill discovered	Load a skill’s full instructions into context. See Skills.
switch_persona	At least one persona defined	Queue a persona switch for the end of the current turn. See Personas.
mcp__*	MCP servers configured	Tools from connected MCP servers. See MCP Servers.

Secret redaction

All tools that return file content or command output run the result through secret redaction before returning. This catches:

• Pattern-based secrets (API keys, tokens, passwords matching common formats)
• File-based secrets (contents of files listed in secrets.files)
• Word-based secrets (specific values listed in secrets.words)

See Permissions for how to configure secret redaction.

Adding a custom tool

See Adding a Tool for the developer guide on implementing and registering new tools.