Shiplight

Execute browser actions using element indices from inspect_page. CRITICAL: Before your first act call, you MUST read resource 'shiplight://schemas/action-entity' to learn action parameter formats. Each action has specific required parameters (e.g., press uses "keys" not "key"). To read it: first call ListMcpResourcesTool (without a server parameter) to discover the correct server name, then call ReadMcpResourceTool with that server name and uri 'shiplight://schemas/action-entity'. Workflow: 1. inspect_page → get DOM with element indices [0], [1], etc. + screenshot 2. act → execute actions using element indices 3. Repeat until task is complete Supported actions: click, double_click, right_click, hover, input_text, clear_input, press, select_dropdown_option, get_dropdown_options, set_date_for_native_date_picker, scroll, scroll_to_text, scroll_on_element, go_to_url, go_back, reload_page, switch_tab, close_tab, wait, wait_for_download_complete, upload_file, save_variable Each action is exactly ONE object with the action name as its single key (e.g. input_text, press). Empty objects {} are invalid and will error. When converting from TestFlow YAML (action_entity with action_data.action_name/kwargs), map each statement to this shape: key = action_name, value = { element_index, description, ...kwargs }. - element_index: Element index from DOM (for element-based actions) - description: REQUIRED - Semantic, human-readable description of the action - Action-specific params: MUST read 'shiplight://schemas/action-entity' resource (use ListMcpResourcesTool without server to discover the server name first) CRITICAL: The 'description' field must be semantic and grounded - DO NOT reference element indices. - BAD: "Click element 5", "Input text to element 3" - GOOD: "Click the Submit button", "Enter '[email protected]' in email field" Response fields: - success: Whether all actions completed successfully - actions: Array of { action_entity, success, error?, screenshot_path? } for each executed action - actions[].screenshot_path: Path to the post-action screenshot (present when screenshot was captured successfully). - dom_file_path: Path to the DOM text file with updated page state and element indices. Absent if the page was still loading — call inspect_page if not present.

Connect to a browser for testing. Two modes: 1. Direct CDP: pass cdp_url (e.g. from a debugger's /api/browser-cdp endpoint) to connect directly via Chrome DevTools Protocol. 2. Extension relay: omit cdp_url to auto-discover Chrome tabs via the Shiplight AI Chrome extension. Returns session_id for use with inspect_page, act, get_page_info, switch_tab, etc.

Close a browser session. Omit session_id entirely to close ALL open sessions.

Generate a self-contained HTML session report. Can be called while the session is live OR after close_session. Embeds per-step screenshots, an optional video player, and a Playwright trace viewer link. The session must have been started with record_evidence: true. Pass local_video_path (returned by close_session) to embed the video. Pass local_trace_path (returned by close_session) to include trace info for later upload. Pass trace_url if the trace has already been uploaded (shows an interactive viewer link). Pass summary to describe what was tested. Pass checks — a list of acceptance criteria derived from what you were asked to build — each with a pass/fail result and the step indices that prove it (step indices are 1-based: first action = step 1). The agent should fill these in itself based on the task and steps taken; no user input required. Pass highlight_steps to feature specific screenshots in a Summary section at the top. Returns the file path of the written report.

Get console logs from the browser (errors, warnings, etc.)

Get network request logs from the browser

Extract Playwright locators for one or more elements by index. Call inspect_page first. Returns locator, xpath, frame path, tag name, and text for each element. Use this to collect locator data for building test flows without interacting with the page. Pass multiple indices in a single call to save round-trips.

Get basic page information (URL and title) without taking a screenshot. Use this as a lightweight way to check what page you're on. For visual inspection, use inspect_page.

Inspect the current page: extracts the DOM tree and saves a Set-of-Mark screenshot. Returns the DOM file path (with element indices [0], [1], etc. for the act tool) and a screenshot file path. The DOM and screenshot are captured atomically so element indices always match. IMPORTANT: Read the DOM file to understand page state and decide actions. Only view the screenshot file when the DOM is insufficient (e.g., visual layout, canvas content, images, verifying colors/styling). Returns: - dom_file_path: Path to file containing DOM text with element indices - screenshot_path: Path to the SoM screenshot (element indices match the DOM) - current_url: The current page URL

Navigate to a URL in an existing browser session. Examples: - Navigate to a specific page: url="https://example.com/login" - Stay at current page and get URL: (no url parameter) The tool returns the current URL after navigation.

Create a new browser session with optional device emulation. Returns a session_id for subsequent operations. Use storage_state_path to restore a previously saved session (cookies, localStorage, IndexedDB). SCREEN RECORDING: Set browser_options.record_evidence=true only when you believe implementation is complete and are doing a final verification pass — not for exploration, investigation, or mid-task checking. This produces a video and trace the user can watch to confirm the finished change works. Both are saved automatically when the session is closed. After creating a session, IMMEDIATELY read resource 'shiplight://schemas/action-entity' for action parameter formats before calling any other tools. To read it: first call ListMcpResourcesTool (without a server parameter) to discover the correct server name, then call ReadMcpResourceTool with that server name and uri 'shiplight://schemas/action-entity'.

Save the browser session's storage state (cookies, localStorage, IndexedDB) to a file. Use this after logging in to cache the session for fast restores via new_session's storage_state_path.