Stop the connection manager and close the connection. Args: timeout: Maximum time to wait for cleanup (default 5s).

Update metrics

Initialize ToolRanker. Args: model_name: Embedding model name. If None, will use env or config value. cache_d

Initialize the wrapper. Args: wrapped_stream: The underlying stderr stream

Add tool result to message history. IMPORTANT: Put screenshot BEFORE text for consistency with initial message.

Hook called after connection is established. Override in subclasses if needed.

Append action to JSONL file.

Add step to traj.jsonl file

Auto-record tool execution to recording manager and quality manager.

Internal method: automatically record tool execution results from LLM client This is called by register_to_llm() wrapper. To

Hook called before establishing connection. Override in subclasses if needed.

Negotiate transport type and set up the appropriate connection manager. Tries transports in order: 1. Streamable HTTP (new MC

Hook called before disconnection. Override in subclasses if needed.

Clean up resources before disconnection.

Build final execution result. Args: instruction: Original instruction messages: Complete conversatio

Build prompt for LLM planning. Args: task_description: The task to accomplish action_history: Previo

Capture screenshot automatically through platform.ScreenshotClient

Check if the npm package is globally installed. Args: package_name: npm package name Returns:

Check if the package is installed. Args: command: The command to check (e.g. "npx", "uvx") args: The argument

Check if the Python package is installed as a uvx tool. uvx tools are installed in ~/.local/share/uv/tools/ directory, not in

Check if local server is available

Check if a Python package is installed via uv pip. Args: package_name: Python package name Returns:

Check workspace directory for existing artifacts that might be relevant to the task. Enhanced to detect if task might already be comp

Hook called when connection fails. Override in subclasses if needed.

Override to add MCP-specific cleanup on connection failure.

Cleanup screenshot client resources

Close the connection. This method should be implemented by subclasses to close the specific type of connection.

Run the connection task. This task establishes and maintains the connection until cancelled.

Create command-line argument parser

Deep merge two dictionaries, update's values will override base's values

Recursively sanitize a JSON schema to conform to JSON Schema draft 2020-12. Removes non-standard fields and ensures valid structure.

Default system prompt for the grounding agent.

Return the screen device number of avfoundation, like '1:none' On macOS, ffmpeg -f avfoundation -list_devices true -i "" wil

Enhance tool result with visual analysis for grounding agent workflows.

Internal implementation of ensure_dependencies (called within lock).

Ensure embedding model is ready (local or remote).

Establish the connection. This method should be implemented by subclasses to establish the specific type of connection needed.

execute the code and track the code block Returns: Tuple[Optional[Dict], str]: (code_info, execution_result)

Execute a planned action through the connector. Args: action: Action dict with action_type and parameters

Execute task with LLM-based planning loop. Planning loop: 1. Observe: Get screenshot 2. Plan: LLM decides ne

Execute LLMClient returned tool_call Args: tool: BaseTool instance (must be obtained from GroundingClient and bound to runtime_info)

Recursively find terminal components

Fix PyAutoGUI '<' character bug.

Fix PyAutoGUI '<' character bug by converting it to hotkey("shift", ',') calls. This fixes the known PyAutoGUI issue where t

Recursively flatten nodes

Format planned action into a human-readable command string. Args: planned_action: Action dictionary from LLM pla

Compact format: one line per step.

Detailed format: multiple lines per step with parameters.

Markdown format: table format.

Retrieve tools with auto-search + cap to control prompt bloat. Falls back to returning all tools if search fails.

Get current cursor image and hotspot Returns: (cursor_image, (hotspot_x, hotspot_y))

Generate default log file path with timestamp (to seconds) Log files are organized by the running script name: - logs/<script

Get the machine architecture, e.g., x86_64, arm64, aarch64, i386, etc. Returns 'amd' for x86/AMD architectures, 'arm' for ARM architectures, or 'u

Get read and write streams from the connection. Override in subclasses if needed.

Get system information for shell agent. First tries to get comprehensive info from local server's /platform endpoint.

Get terminal size

Get workspace directory path from context.

Fallback implementation for PUT/PATCH/DELETE returning JSON/text, can be overridden by subclasses.

Hide cursor

Initialize local fastembed model.

Initialize tool quality manager based on config.

Initialize remote embedding API (OpenRouter/OpenAI compatible).

Execute the install command. Args: command: The command to execute (e.g. "npx", "uvx") args: The arguments li

Validate image response using magic bytes. Args: content_type: HTTP Content-Type header data: Respon

Validate image response using magic bytes.

Internal helper for lazy session creation. Args: server: Server name to create session for Raises:

Lazy import provider class

List tools by starting all servers. Uses a semaphore to serialize session creation, avoiding TaskGroup race conditions that o

Live update loop

LLM pre-filter for MCP servers. Returns (utility_tools, domain_tools, llm_filter_info).

Load configuration

Load log_level from config_grounding.json and convert to ANYTOOL_DEBUG value. Returns: 0 (WARNING), 1 (INFO), or 2 (DEBUG)

Load configuration from multiple files

Load embeddings from disk cache.

Load tools from cache file without starting servers. Priority: 1. Try to load from sanitized cache (mcp_tool_cache_sanitized.

Log search results in a concise, grouped format. Shows backend/server breakdown and tool names (truncated if too many).

Create a shim that wraps streamablehttp_client with improved error handling.

Trigger quality evolution based on global execution count.

Main monitoring loop Args: poll_interval: Update interval in seconds

Get next JSON-RPC request ID.

Parse Anthropic computer tool use to pyautogui command. Args: tool_input: Tool input from Anthropic (action, coo

Parse LLM response to extract action. Args: response: LLM response (should be JSON) Returns:

Use LLM/VLM to plan the next action. This method sends: - Task description - Current screenshot (vision inpu

Convert BaseTool list to LLMClient usable format, with deduplication. Args: tools: BaseTool instance list (should be obtained from Gr

Print system status

Continuously receive and process messages from the WebSocket.

Record agent execution to recording manager. Args: result: Execution result instruction: Original in