Skip to content

Tools reference ​

Every built-in tool the model can call. 31 of them — 20 statically registered, 11 constructed per session (they need a client, model tier, subagent list injected). MCP server tools are added dynamically on top.

The "Permission" column has three values:

  • No — never prompts
  • Depends — on the arguments (is the path inside the working directory, does the command match a rule) and goes through the full permission chain
  • Tools declaring isReadOnly are short-circuited by the permission chain, but deny rules still win

Files ​

ToolRead-onlyPermissionWhat it does
Read✓NoRead a file with line numbers. Use offset/limit for large ones. Any file must be Read before editing
Glob✓NoFind files by glob, max 100, node_modules/.git ignored automatically
Grep✓NoRegex search inside files, returns file:line:content, max 100. ripgrep syntax by default
EditDependsExact string replacement. old_string must match character for character (indentation and newlines included) and be unique by default
WriteDependsWhole-file write, parent directories created. Overwriting requires having Read the file first; new files don't
NotebookEditDependsEdit a single .ipynb cell: replace / insert / delete

Glob for filenames, Grep for contents

Don't shell out to find / grep — the dedicated tools have result caps and ignore rules, so their output costs far less context.

Execution ​

ToolRead-onlyPermissionWhat it does
BashDependsRun a command in a persistent working directory (cd affects every later command). 120s default timeout; output over 30000 chars is truncated in the middle

Network ​

ToolRead-onlyPermissionWhat it does
WebFetchDependsFetch an http(s) URL and extract/summarize per a prompt
WebSearchDependsSearch the web for current information; returns title / link / snippet

Task list ​

ToolRead-onlyPermissionWhat it does
TaskCreateNoCreate a task. Start anything with 3+ steps by listing the plan
TaskUpdateNoUpdate a task. Mark one completed and the next in_progress — at most one in_progress at a time
TaskGet✓NoFetch all fields of a task by id
TaskList✓NoList the current task list

Subagents & orchestration ​

ToolRead-onlyPermissionWhat it does
Agent✓NoDispatch a subagent. See Subagents
Workflow✓DependsOrchestrate subagents with a deterministic JavaScript script. See Workflows
Skill✓NoInvoke a skill. See Skills

Background tasks & scheduling ​

ToolRead-onlyPermissionWhat it does
Monitor✓NoStart a background monitor streaming events from a long-running script. Each stdout line is one event
BgTaskList✓NoList background tasks (id / status / description)
TaskOutput✓NoFetch a background task's output
TaskStop✓NoStop a running background task by id (Monitor, background Bash, cron)
Sleep✓NoWait N seconds; the user can interrupt at any time
CronCreate✓NoQueue a prompt for the future — recurring on a cron schedule, or one-shot
CronList✓NoList cron tasks scheduled in this session
CronDelete✓NoCancel a cron task
ScheduleWakeup✓NoSchedule the next iteration in /loop dynamic mode

Memory ​

ToolRead-onlyPermissionWhat it does
SearchMemory✓NoFull-text search across project memory and the global drawer. See Memory

Session & environment ​

ToolRead-onlyPermissionWhat it does
ConfigDependsRead/write user-level config. Omit value to read the merged value; supply it to write the user layer
ExitPlanMode✓NoSubmit a plan for approval in plan mode. Only available in plan mode
EnterWorktreeNoCreate an isolated git worktree and move the session into it
ExitWorktreeNoLeave the worktree; action=keep or remove
PushNotification✓NoSend a desktop notification to pull attention back. This has a cost — prefer not to
AskUserQuestion✓NoPop a structured question when the user should decide, instead of guessing. 1–4 questions, 2–4 options each

Denied to subagents ​

Regardless of type, these 11 are denied to every subagent:

ExitPlanMode  EnterWorktree  ExitWorktree  Workflow
ScheduleWakeup  CronCreate  CronList  CronDelete
Monitor  TaskStop  PushNotification

They either mutate session-level state or spawn more work — a subagent must not open branches of its own. Individual types add further restrictions (Explore / Plan / verification all deny Edit/Write); see Subagents.

MCP tools ​

Tools provided by MCP servers are injected at runtime and namespaced by server. /mcp shows connected servers and what they expose. See MCP.


Related: Tools overview · Permissions · Subagents