click an element on the page

close the browser. args: {}

read the page's JS console output and exceptions. args: {}

evaluate JavaScript in the page context

send keyboard keys to the page

open a URL (or local file / dev server). Returns title + page text + interactive elements. args: { "url": string }

all visible text of the current page + element list (incl. current input values); a selector reads only its matches. args: { "selector"? }

reload the current page (cache ignored); required after local code edits — without it you see the stale page. args: {}

FULL-page shot (tall pages auto-segment; heavy). First look only — re-checks: browser_snapshot. args: {}

scroll to load more. args: { "to"?: "bottom"|"top", "by"?: number }

viewport shot (instant, light). PREFER for re-checks and after scrolling. args: {}

type text into a focused element

exact replacement (old_string must match the file verbatim and be unique unless replace_all=true). For several changes in one file, pass an atomic edits array in ONE call (any failure = nothing applied) instead of many small calls. args: { "path", "old_string", "new_string", "replace_all"? } or { "path", "edits": [{ "old_string", "new_string", "replace_all"? }] }

edit by line anchors (hidden from doc, anchor mode only)

find files by pattern (e.g. "src/**/*.ts"). args: { "pattern": string }

regex search over file contents. args: { "pattern": string, "path"?: string, "glob"?: string }

list one directory level (no path = workspace root). args: { "path"?: string }

atomic multi-edit (undocumented, executable)

read several files in one call — reach for it instead of firing read_file again and again. A file that cannot be read is reported on its own; the rest still come back. args: { "paths": string[] }

definition outline (functions/classes + line numbers) without reading the whole file. args: { "path": string }

read a file (pdf/docx/xlsx/pptx too — text auto-extracted, scans OCR'd). Pass symbol to get one function/class definition plus its call sites. args: { "path": string, "offset"?: number(1-based), "limit"?: number, "symbol"?: string }

ask the codebase by meaning; returns relevance-ranked files with their key definitions. First choice for unfamiliar code. args: { "query": "where login auth is handled", "k"?: number }

search the user's knowledge-base documents. args: { "query": string }

literal keyword search over file names + contents; names_only=true for names only. args: { "query": string, "path"?: string, "names_only"?: boolean }

one-call repo overview (README, manifest, tree, languages, entry points). First move in an unfamiliar workspace. args: {}

create/update the todo plan; mark finished steps done, the next one in_progress. args: { "todos": [{ "content": string, "status": "pending"|"in_progress"|"done" }] }

find and run just the tests related to the change; no args = everything changed this turn. args: { "files"?: string[] }

view an image in the workspace (vision models see it; others get OCR'd text). args: { "path": string }

background-download a file into the workspace; you're notified on completion — don't read it or re-request before that. args: { "url": string, "path": string }

fetch a URL, handled by content type (article→Markdown, GitHub file→raw source, video→transcript, PDF→text); raw=true for HTML. args: { "url": string, "raw"?: boolean }

web search; site scopes to one site (github.com returns structured repos/issues/code; reddit/youtube/bilibili/any domain). args: { "query": string, "site"?: string }

create a file, or rewrite one wholesale (replaces ALL content); to modify an existing file prefer edit_file. args: { "path": string, "content": string }