mirror of
https://github.com/mcp-use/mcp-use.git
synced 2026-09-14 18:06:01 +08:00
Luigi Pederzani
f08e870e09
* chore: update imports * refactor(mcpagent): update agent creation logic and imports - Replaced `create_tool_calling_agent` with `create_agent` for improved agent instantiation. - Simplified the `_create_agent` method by removing the `AgentExecutor` and directly returning the created agent. - Adjusted type annotations for `_agent_executor` to initialize as `None` instead of using a union type. * feat(mcpagent): migrate to LangChain 1.0.0 and simplify agent methods - Updated MCPAgent to utilize LangChain 1.0.0's `create_agent()` for improved agent instantiation. - Introduced new methods `astream_simplified()` and `run_v2()` that leverage the built-in `astream()` for internal agent loop handling. - Retained legacy methods `stream()` and `run()` for backward compatibility with manual execution. - Enhanced documentation to reflect changes and provide clearer usage examples. * feat(mcpagent): implement dynamic tool updates during agent execution - Enhanced the MCPAgent to support dynamic tool updates when using server_manager mode. - Implemented logic to check for tool changes before and during execution, allowing for immediate updates and safe restarts. - Introduced a maximum restart limit to prevent infinite loops during tool updates. - Updated documentation to reflect the new behavior and usage of dynamic tool management. * refactor(mcpagent): clean up imports and streamline logging - Removed unused imports and organized existing ones for better clarity in `mcpagent.py`. - Simplified logging statements for consistency and improved readability. - Enhanced the overall structure of the code by eliminating unnecessary whitespace. - Updated unit tests to reflect changes in imports and ensure compatibility with the refactored code. * feat(mcpagent): enhance agent execution with step tracking and middleware - Introduced `ModelCallLimitMiddleware` to enforce a maximum number of model calls during agent execution. - Updated the `_consume_and_return` method to return a tuple of the final result and the number of steps taken. - Enhanced telemetry tracking to include steps taken and execution time for better performance insights. - Simplified agent creation logic by integrating middleware directly into the `create_agent` function. - Improved logging for agent execution details, including steps and tool usage. * fix(mcpagent): add null check for node output messages extraction - Updated the MCPAgent to include a null check for `node_output` before attempting to access the "messages" key, preventing potential errors during message extraction. * minor * feat(mcpagent): filter out ToolMessage from history - Updated the MCPAgent to exclude ToolMessage instances from being added to the message history, ensuring that only relevant messages are recorded. - This change helps maintain a cleaner history by avoiding intermediate tool execution details. * fix(mcpagent): set max_steps assignment * ruff * docs * e2e tests for agent * key in tests * fix manager test and split ci tests * telemetry off for tests * fix server manager test * feat(mcpagent): delegate query streaming to remote agent in remote mode - recreated functionality to the MCPAgent to yield results from a remote agent when operating in remote mode, enhancing the agent's capability to handle queries efficiently. * fix: agent unit tests * Print inputs * rollback, wrong repo * fix streaming output * stricter streaming test * pass observability callbacks * fix tests mock --------- Co-authored-by: pietrozullo <pietro.zullo@gmail.com>