Tools reference

Tools are functions the model can call during a task. Which tools are available depends on the prompt profile (Standard or Lite), the current mode, feature toggles, provider capabilities, and task permissions. The “Conditions” column lists only the requirements specific to each tool; every tool is also subject to your approval settings.

Files and code

Tool	Purpose	Conditions
`read_file`	Read file content, optionally a line range	Standard, Lite
`write_to_file`	Create a new file, creating directories as needed	Standard, Lite
`replace_in_file`	Edit an existing file with line-matched SEARCH/REPLACE blocks	Standard, Lite
`search_files`	Search file contents in a directory with a regular expression	Standard, Lite
`list_files`	List directory contents, optionally recursively	Standard, Lite
`list_code_definition_names`	List top-level source-code definitions	Standard, Lite
`apply_patch`	Add, delete, move, or edit code using a patch format	Standard only
`find_references`	Find symbol references through the IDE language service	Standard only; requires VS Code
`rename`	Rename a symbol through the IDE language service, with preview	Standard only; requires VS Code
`replace_text`	Replace text across files matching a glob, with preview	Standard only

How `replace_in_file` matches

Each SEARCH line matches a whole file line or the beginning of one. Leading indentation may be omitted, and a SEARCH line never matches from the middle of a file line.
A SEARCH block must match exactly one location in the file. A block that matches several locations is rejected, and the result lists the candidate line numbers.
Every matched line is replaced in full. REPLACE is written verbatim, so replacement lines must include their indentation.
A SEARCH line of the form ....... SKIP (as many dots as the delimiter length) marks a SKIP range. The block then covers every line from the head through the nearest matching tail, and all of those lines are replaced by REPLACE.
A successful edit reports the original line range each block replaced, including how many lines a SKIP range removed.

replace_text is a separate tool and is not affected by these rules.

Commands, browser, and web

Tool	Purpose	Conditions
`execute_command`	Run a shell command	Standard, Lite
`kill_command`	Stop a command started by `execute_command`	Standard, Lite
`browser_action`	Operate a Puppeteer-controlled browser	Standard only; the model supports browser use and it is not disabled in browser settings
`web_fetch`	Fetch and analyze a given URL	Standard only; Web Tools is on
`web_search`	Search the web and return titles and links	Standard only; Web Tools is on
`generate_image`	Generate or edit images through the configured image source	Enable Image Generation is on and an image model is available

Task interaction

Tool	Purpose	Conditions
`ask_followup_question`	Ask the user one question that needs an answer	Standard, Lite
`qna_respond`	Answer the user’s question	Standard, Lite
`make_plan`	Submit a complete implementation or design plan	Standard, Lite
`working_respond`	Tell the user what the next tool call does and why, reply to a queued or mid-task message, or acknowledge a correction, then keep working. Never ends the turn and must be followed by a work tool. Same tool in Plan and Act mode; calls under the old name `act_mode_respond` still run it	Standard, Lite
`phase_report`	Report a finished phase: what was completed, the verification result, and the next phase. Never ends the turn; with `requires_acknowledgment` it waits for the user to acknowledge or stop. Available in Plan and Act mode; calls under the old name `status_update` still run it	Standard, Lite
`generate_report`	Submit a structured report for review	Standard, Lite
`attempt_completion`	Submit the final result of the task	Standard, Lite
`change_todo_list`	Ask the user to approve replacing the current to-do list	Standard only; Focus Chain is on

Capabilities and delegation

Tool	Purpose	Conditions
`load_skill`	Load the full instructions of a skill	Standard only; at least one skill is available
`load_workflow`	Load the steps of a workflow	Standard only
`load_mcp`	Load the metadata of one MCP tool without running it	Standard only; an MCP server is connected
`use_mcp_tool`	Run a tool provided by an MCP server	Standard only; an MCP server is connected
`access_mcp_resource`	Read a resource provided by an MCP server	Standard only; an MCP server is connected
`load_mcp_documentation`	Load guidance for creating and installing MCP servers	Standard only; an MCP server is connected
`use_subagent` / `use_subagents`	Run one subagent or a batch of subagents	Standard only; subagents are enabled and the current run is not itself a subagent
`spawn_task`	Start an independent new task; requires user approval before creation	Standard only; not inside a subagent or a spawned task

Strict Plan Mode

When Strict Plan Mode is on in Settings → Features, Plan mode rejects write_to_file, replace_in_file, and apply_patch, as well as the rule-creation tool used by /newrule.

MCP tools

Besides the built-in tools above, tools provided by connected MCP servers also become callable. MCP server configuration comes from the global file <documents-root>/settings/mcp_settings.json and the workspace directory .agents/mcp/. See MCP.