Tools Reference

The agent comes equipped with a comprehensive suite of tools. When you ask it to perform a task, it autonomously selects and chains these tools to get the job done.

Tab handles

Every tab-interacting tool takes an explicit tab argument — a stable handle (t1, t2, ...) that identifies which tab the call should act on. Handles are minted by navigate (when it opens a new tab) and selectTab (when binding an external tab into the conversation). Once minted, a handle keeps referring to the same tab on subsequent turns and across service-worker restarts.

The agent sees the live set of available handles in a ## Tabs in this conversation block injected into its system prompt every turn. Calling listTabs returns the same shape and is the way to discover tabs the user has open elsewhere.

To bootstrap a fresh conversation, the agent calls navigate({ url }) with no tab arg — that opens a new background tab and returns the new handle in the response.

Browsing & Interaction

The agent navigates the web and interacts with elements just like a human user. Each tool below requires tab (except navigate, where omitting tab opens a new tab).

Tab Management

Extraction

Batching

Batching collapses several round-trips into one. It is most useful when the agent already knows every argument up front — reading three tabs, grepping two patterns, or firing several webSearch queries.

batch({
  description: "Comparing pricing pages",
  invocations: [
    { name: "snapshot", arguments: { tab: "t1", mode: "viewport" } },
    { name: "readPage", arguments: { tab: "t2" } },
    { name: "webSearch", arguments: { query: "competitor pricing" } }
  ]
})

description is a short present-participle phrase naming the work in the user's terms — it becomes the label on the tool call in chat. The agent writes the phrase; the interface adds progress and outcome. A batch labelled Comparing pricing pages reads as "Comparing pricing pages… 4 reads" while it runs, then "Comparing pricing pages · 4 of 4" when every call succeeds, or "Comparing pricing pages · 3 of 4" with an amber alert when one fails. The phrase never changes tense, so it stays accurate regardless of the language the agent is writing in.

Batchable tools: snapshot, readPage, extract, read_network_requests, read_console_messages, listTabs, webSearch, Read, Glob, Grep, LS, searchMemory, skill, list_artifacts, read_artifact_diagnostics.

Invocations run concurrently and cannot see each other’s output, so calls whose arguments depend on a previous result — an @ref from a snapshot, a URL from a search — must be issued separately. Subagents get a batch scoped to their own allowed tools, so batching can never widen a subagent’s reach.

Code Execution

The agent can execute code in three sandboxed environments depending on what it's doing.

For Python specifically, the agent loads the bundled python-env skill on demand to pick up Pyodide-specific gotchas (no subprocess, mismatched import names like import fitz for pymupdf, top-level await, latin-1 PDF unicode pitfalls, etc.).

Filesystem

These tools operate on the conversation's per-conversation virtual Workspace, backed by OPFS.

Skills

Skills let the agent load curated, on-demand instructions instead of carrying everything in its system prompt. See Skills for the full picture.

To read a supporting file bundled with a skill (reference docs, scripts), the agent uses the Read tool with the skill's path, e.g. Read({ file_path: "/skills/<name>/references/<file>" }).

Task Planning

For complex, multi-step operations, the agent maintains a structured plan visible in the side panel's Cowork → Progress card.

Asking You

The agent works autonomously and picks a reasonable default rather than checking in. askUser is the narrow exception: when a decision is genuinely yours and getting it wrong is expensive or irreversible — which flight to book, which saved card to pay with, which account to post from — it pauses and asks.

The question replaces the chat composer, so answering is the only thing in front of you. Questions are shown one at a time — move between them with the arrow keys, the back/forward arrows, or the progress dots in the header. Pick options with the number keys, type a custom answer in the always-present free-text box (which takes multiple lines; Shift+Enter for a newline, Esc to hand the arrow keys back), or choose Skip all questions to hand the whole call back to the agent — that discards anything you'd already answered and it proceeds on its best interpretation, telling you what it assumed. To skip just one question, leave it blank and move on: a blank question is reported as unanswered rather than as an empty answer.

The agent is genuinely stopped while a question is open: askUser runs in your browser rather than in the agent loop, so the turn ends and a new one starts when you answer. That means a question survives a reload, a side-panel close, and a browser restart — and there's no deadline to answer it. Nothing is held open while it waits, so you can leave a question sitting and answer it tomorrow; the agent picks up from your answer whenever it arrives.

Questions are unavailable where there's nobody to ask: scheduled tasks, runs driven by an external MCP host, and subagents never get the tool.

Memory

Persistent memory across conversations. See Memory for details.

Connectors (MCP)

Any Connector you enable adds its own MCP tools to the agent's toolset dynamically. Names are namespaced by connector (e.g. github_search_issues, linear_create_issue).