Tools of the model

The model calls these tools with your permissions; Tatu has no per-tool permission gates.

Tool	What it does	Main parameters	See
bash	Run a shell command in the project; each call starts a fresh shell with no stdin. Background execution is available only in interactive sessions.	Required command; optional timeout in seconds, run_in_background when available	Background tasks
read	Read text or attach an image; long text can be read in ranges. A path that does not exist is answered with close names or with what the nearest folder holds.	Required path; optional offset (1-based line), limit	Images
write	Create a file and parent folders, or overwrite a file already read and unchanged since.	Required path, content (file_text is accepted for it)	Security and trust
edit	Replace exact text in a file already read; all replacements must match and not overlap.	Required path, edits; each edit has oldText, newText, optional replaceAll. An oldText and newText beside edits count as one more edit	Security and trust
grep	Search file contents with a Go regular expression or literal text; respect .gitignore and skip binary files.	Required pattern; optional path, glob, ignoreCase, literal, context, limit	Your first session
find	Find paths by glob, respecting .gitignore.	Required pattern; optional path, limit	Your first session
web_search	Search third-party services for titles, URLs and snippets; queries leave your computer.	Required query; optional limit (default 8, maximum 20)	The web
web_fetch	Read a page as Markdown; save large pages locally or ask a small model a question about the page. Private addresses are never sent to third-party readers.	Required url; optional prompt	The web
agent	Give a subagent a task and receive its report; it has its own conversation and cannot start another subagent or ask you questions.	Required description, prompt; optional type, model (fast, balanced, main), step	Subagents
ask_user	Ask you 1–4 multiple-choice questions, with 2–5 options each and a write-your-own-answer row; interactive sessions only.	Required questions; each has question, options with label; optional header, multiSelect, option description and preview	Your first session
preview	Share a local web app with the Tatu app or a paired remote terminal: start runs the dev server as a background task, waits for its port, checks the HTTP response at / and publishes it; publish shares a server that already runs (it fails when nothing listens); revoke and list. Interactive main agent only, not subagents. Publication is explicit; starting a server separately grants no access. It does not validate rendering or monitor server health.	Required action (start, publish, revoke, list); start requires command, optional cwd, port, extra_ports, name; publish requires port, optional extra_ports (up to three other servers the page calls), name; revoke requires id (publication ID or all); list takes no other parameters	The app
plan_write	Replace the plan with a goal and 2–15 steps; keep the status of retained step IDs.	Required goal, steps; step id, title, with optional detail, files, dependsOn, verify, parallel, parallelWhy, status; optional askApproval (ask you to approve the plan, as /plan does) and approval (the words of that question)	Planning
plan_update	Change step statuses in order, enforcing dependencies and parallel-work rules.	Required updates; each has id, status (pending, in_progress, completed, cancelled), optional note	Planning
task_output	Inspect a background task and the end of its output, optionally waiting for it or for a line of its output; interactive sessions only.	Required task_id; optional block, timeout (default 30 seconds, maximum 600), until (with block: a regular expression; return as soon as the output matches, such as a server’s ready line)	Background tasks
task_stop	Stop a background task and its child processes; interactive sessions only.	Required task_id	Background tasks
codemode	Run a JavaScript program in an isolated sandbox that calls the other tools (tools.<name>(args)), so many calls and their intermediate results cost one step; only what the program prints comes back. Calls go through your hooks like the model’s own, and a program can keep small values between runs with store and load. Turn it off with codemode.enabled.	Required code; an optional first line // @options: {"max_output_tokens": 10000, "timeout_ms": 300000}	Programs
mcp__<server>__<tool>	A tool of an MCP server you configured; one entry per tool.	The server’s own parameters	MCP servers