Tools

What makes Mini-claude an agent and not a chat: it calls tools to act. The model decides which one; Mini-claude runs it on your machine and feeds the result back.

Files

read_file: reads a file in the workspace.
write_file: creates or replaces a file.
edit_file: modifies an existing file via targeted replacement.
grep / glob: search text or files by pattern.
tree: prints the project tree.

write_file and edit_file change state: they ask permission and show a diff before acting. See Permissions & safety.

Commands

bash: runs a shell command in the workspace.

Like anything that touches state, every command goes through the permission prompt.

Web & browser: from your machine

This is where Mini-claude stands out. It outsources nothing to a remote service: it drives your own Chrome, from your machine, with your IP.

No account to share, no remote session, no “let this AI log into your account.” The agent browses the way you would, on your own box. That’s more private and safer.

Visible browser (you watch the tab live):

browser_open: open a URL.
browser_click: click an element by its visible text.
browser_type: fill a field (and optionally submit with Enter).
browser_key: press a key.
browser_scroll: scroll the page.
browser_fill: fill a whole form.
browser_read: read the current page.
browser_screenshot: capture the page.

Without opening a tab (to answer a factual question):

web_search: search to discover unknown pages.
web_fetch: silently fetch the text of a known URL.

The rule: when you want to see or interact with a site, Mini-claude opens the browser. When it just needs text to answer, it fetches silently.

Availability per mode

Not all tools are active everywhere. guided has no tools (it only asks questions). review is read-only (read files + web, no writes). auto and explain have the full set. See The 4 modes.

The 4 modes Permissions & safety