stealth-browser-mcp

Stealth Browser MCP

Core Workflow

1. Start with spawn_browser and keep the returned instance_id for every follow-up tool call.
2. Use navigate for page loads, then verify state with get_instance_state, get_page_content, or take_screenshot.
3. Inspect before acting. Use query_elements, get_page_content, and screenshots to find stable selectors and confirm the target element is visible.
4. Act with the narrowest tool that fits the task, such as click_element, type_text, scroll_page, wait_for_element, or file_upload.
5. Verify after each important action with page content, DOM queries, URL/title state, or a screenshot.
6. Close browsers with close_instance when the task is complete unless the user explicitly wants the session left open.

Tool Selection

• Use --minimal when an agent only needs browser spawning, navigation, screenshots, page content, and basic interaction.
• Use full mode for CDP execution, network debugging, cookie/storage work, dynamic hooks, and advanced element extraction.
• Use execute_script for normal page JavaScript.
• Use execute_cdp_command only when a CDP primitive is required or when page JavaScript cannot access the needed browser state.
• Use network tools after navigation starts: list_network_requests, search_network_requests, get_request_details, get_response_details, and get_response_content.
• Use extraction tools when the task is to recreate, clone, or analyze page elements with styles, assets, events, or animations.

Pre-Document Scripts

Use add_script_to_evaluate_on_new_document before navigation when a site must see modified browser APIs during its first inline scripts. This is required for WebGL renderer/vendor spoofing, navigator property patches, and similar anti-bot checks that run during page parse.

Keep world_name=None for main-world browser API spoofing. Set run_immediately=True when the patch should also run in the current document.

Reliability Rules

• Prefer stable CSS selectors, roles, visible text, or attributes over brittle generated classes.
• Wait for page state instead of assuming navigation or rendering has finished.
• Treat every tool result as stateful: pass the exact instance_id, request IDs, and selectors returned by prior steps.
• For file uploads, pass absolute file paths inside directories allowed by BROWSER_FILE_UPLOAD_ALLOWED_DIRS.
• If output is too large, ask for a narrower selector, use filters, or rely on file-backed extraction responses.
• Keep debug logging disabled during normal MCP stdio usage. Enable --debug or STEALTH_BROWSER_DEBUG=1 only when diagnosing server behavior.