🐫 CAMEL: The first and the best multi-agent framework. Finding the Scaling Law of Agents. https://www.camel-ai.org
Python agentai-societiesartificial-intelligencecommunicative-aicooperative-aideep-learninglarge-language-modelsmulti-agent-systemsnatural-language-processing
feat: update browser_nonvisual_human_in_the_loop and EdgeOnePagesMCPToolkit (#2710)
Tao Suna3c30db8844e
diff --git a/camel/toolkits/__init__.py b/camel/toolkits/__init__.pyindex d1e5354af..ab929931c 100644--- a/camel/toolkits/__init__.py+++ b/camel/toolkits/__init__.py@@ -78,6 +78,7 @@ from .playwright_mcp_toolkit import PlaywrightMCPToolkitfrom .wolfram_alpha_toolkit import WolframAlphaToolkitfrom .task_planning_toolkit import TaskPlanningToolkitfrom .non_visual_browser_toolkit import BrowserNonVisualToolkit+from .edgeone_pages_mcp_toolkit import EdgeOnePagesMCPToolkit__all__ = [@@ -144,4 +145,5 @@ __all__ = ['BohriumToolkit','TaskPlanningToolkit','BrowserNonVisualToolkit',+ 'EdgeOnePagesMCPToolkit',]diff --git a/camel/toolkits/edgeone_pages_mcp_toolkit.py b/camel/toolkits/edgeone_pages_mcp_toolkit.pynew file mode 100644index 000000000..f17c02ffd--- /dev/null+++ b/camel/toolkits/edgeone_pages_mcp_toolkit.py@@ -0,0 +1,69 @@+# ========= Copyright 2023-2024 @ CAMEL-AI.org. All Rights Reserved. =========+# Licensed under the Apache License, Version 2.0 (the "License");+# you may not use this file except in compliance with the License.+# You may obtain a copy of the License at+#+# http://www.apache.org/licenses/LICENSE-2.0+#+# Unless required by applicable law or agreed to in writing, software+# distributed under the License is distributed on an "AS IS" BASIS,+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.+# See the License for the specific language governing permissions and+# limitations under the License.+# ========= Copyright 2023-2024 @ CAMEL-AI.org. All Rights Reserved. =========++from typing import List, Optional++from camel.toolkits import BaseToolkit, FunctionTool++from .mcp_toolkit import MCPToolkit+++class EdgeOnePagesMCPToolkit(BaseToolkit):+ r"""EdgeOnePagesMCPToolkit provides an interface for interacting with+ EdgeOne pages using the EdgeOne Pages MCP server.++ Attributes:+ timeout (Optional[float]): Connection timeout in seconds.+ (default: :obj:`None`)+ """++ def __init__(+ self,+ timeout: Optional[float] = None,+ ) -> None:+ r"""Initializes the EdgeOnePagesMCPToolkit.++ Args:+ timeout (Optional[float]): Connection timeout in seconds.+ (default: :obj:`None`)+ """+ super().__init__(timeout=timeout)++ self._mcp_toolkit = MCPToolkit(+ config_dict={+ "mcpServers": {+ "edgeone-pages-mcp-server": {+ "command": "npx",+ "args": ["edgeone-pages-mcp"],+ }+ }+ },+ timeout=timeout,+ )++ async def connect(self):+ r"""Explicitly connect to the EdgeOne Pages MCP server."""+ await self._mcp_toolkit.connect()++ async def disconnect(self):+ r"""Explicitly disconnect from the EdgeOne Pages MCP server."""+ await self._mcp_toolkit.disconnect()++ def get_tools(self) -> List[FunctionTool]:+ r"""Returns a list of tools provided by the EdgeOnePagesMCPToolkit.++ Returns:+ List[FunctionTool]: List of available tools.+ """+ return self._mcp_toolkit.get_tools()diff --git a/camel/toolkits/non_visual_browser_toolkit/browser_non_visual_toolkit.py b/camel/toolkits/non_visual_browser_toolkit/browser_non_visual_toolkit.pyindex cae8dd274..a4d1b9129 100644--- a/camel/toolkits/non_visual_browser_toolkit/browser_non_visual_toolkit.py+++ b/camel/toolkits/non_visual_browser_toolkit/browser_non_visual_toolkit.py@@ -115,16 +115,19 @@ class BrowserNonVisualToolkit(BaseToolkit):async def open_browser(self, start_url: Optional[str] = None) -> Dict[str, str]:- r"""Launch a Playwright browser session.+ r"""Launches a new browser session. This should be the first step.Args:- start_url (Optional[str]): Optional URL to navigate to immediately- after the browser launches. If not provided, the browser will- not navigate to any URL. (default: :obj:`None`)+ start_url (Optional[str]): The URL to navigate to after the browser+ launches. If not provided, the browser will open with a blank+ page. (default: :obj:`None`)Returns:- Dict[str, str]: Keys: ``result`` for action outcome,- ``snapshot`` for full DOM snapshot.+ Dict[str, str]: A dictionary containing the result of the action+ and a snapshot of the page. The keys are "result" and+ "snapshot". The "snapshot" is a YAML-like representation of+ the page's DOM structure, including element references+ (e.g., "e3") that can be used in other tool calls."""await self._session.ensure_browser()if start_url:@@ -136,10 +139,12 @@ class BrowserNonVisualToolkit(BaseToolkit):return {"result": "Browser session started.", "snapshot": snapshot}async def close_browser(self) -> str:- r"""Terminate the current browser session and free all resources.+ r"""Closes the current browser session, freeing up all associated+ resources. This should be called when the browsing task is complete.Returns:- str: Confirmation message.+ str: A confirmation message indicating the session has been+ closed."""# Close agent if it existsif self._agent is not None:@@ -154,14 +159,17 @@ class BrowserNonVisualToolkit(BaseToolkit):return "Browser session closed."async def visit_page(self, url: str) -> Dict[str, str]:- r"""Navigate the current page to the specified URL.+ r"""Navigates the current browser page to a new URL.Args:- url (str): The destination URL.+ url (str): The URL to navigate to. Must be a fully qualified URL+ (e.g., "https://www.google.com").Returns:- Dict[str, str]: Keys: ``result`` for action outcome,- ``snapshot`` for full DOM snapshot.+ Dict[str, str]: A dictionary containing the navigation result and a+ snapshot of the new page. The keys are "result" and+ "snapshot". The "snapshot" provides a fresh view of the+ page's DOM structure."""if not url or not isinstance(url, str):raise ValueError("visit_page(): 'url' must be a non-empty string")@@ -192,14 +200,16 @@ class BrowserNonVisualToolkit(BaseToolkit):)async def click(self, *, ref: str) -> Dict[str, str]:- r"""Click an element identified by ``ref``+ r"""Performs a click action on a specified element on the current page.Args:- ref (str): Element reference ID extracted from snapshot- (e.g.``"e3"``).+ ref (str): The reference ID of the element to click. This ID is+ obtained from the page snapshot (e.g., "e12").Returns:- Dict[str, str]: Result message from ``ActionExecutor``.+ Dict[str, str]: A dictionary containing the result of the action.+ If the click causes a change in the page's structure, a+ "snapshot" key will be included with a diff of the changes."""self._validate_ref(ref, "click")@@ -207,15 +217,16 @@ class BrowserNonVisualToolkit(BaseToolkit):return await self._exec_with_snapshot(action)async def type(self, *, ref: str, text: str) -> Dict[str, str]:- r"""Type text into an input or textarea element.+ r"""Types text into an input field or textarea on the current page.Args:- ref (str): Element reference ID extracted from snapshot.- (e.g.``"e3"``).- text (str): The text to enter.+ ref (str): The reference ID of the input element. This ID is+ obtained from the page snapshot (e.g., "e25").+ text (str): The text to be typed into the element.Returns:- Dict[str, str]: Execution result message.+ Dict[str, str]: A dictionary containing the result of the action.+ The key is "result"."""self._validate_ref(ref, "type")@@ -223,14 +234,17 @@ class BrowserNonVisualToolkit(BaseToolkit):return await self._exec_with_snapshot(action)async def select(self, *, ref: str, value: str) -> Dict[str, str]:- r"""Select an option in a ``<select>`` element.+ r"""Selects an option from a dropdown (<select>) element on the page.Args:- ref (str): Element reference ID.- value (str): The value / option to select.+ ref (str): The reference ID of the <select> element. This ID is+ obtained from the page snapshot.+ value (str): The value of the option to be selected. This should+ match the 'value' attribute of an <option> tag.Returns:- Dict[str, str]: Execution result message.+ Dict[str, str]: A dictionary containing the result of the action.+ The key is "result"."""self._validate_ref(ref, "select")@@ -238,15 +252,16 @@ class BrowserNonVisualToolkit(BaseToolkit):return await self._exec_with_snapshot(action)async def scroll(self, *, direction: str, amount: int) -> Dict[str, str]:- r"""Scroll the page.+ r"""Scrolls the current page up or down by a specified amount.Args:- direction (str): Scroll direction, should be ``"down"`` or- ``"up"``.- amount (int): Pixel distance to scroll.+ direction (str): The direction to scroll. Must be either "up" or+ "down".+ amount (int): The number of pixels to scroll.Returns:- Dict[str, str]: Execution result message.+ Dict[str, str]: A dictionary containing the result of the action.+ The key is "result"."""if direction not in ("up", "down"):logger.error("scroll(): 'direction' must be 'up' or 'down'")@@ -258,53 +273,72 @@ class BrowserNonVisualToolkit(BaseToolkit):return await self._exec_with_snapshot(action)async def enter(self, *, ref: str) -> Dict[str, str]:- r"""Press the Enter key.+ r"""Simulates pressing the Enter key on a specific element.+ This is often used to submit forms.Args:- ref (str): Element reference ID to focus before pressing.+ ref (str): The reference ID of the element to focus before+ pressing Enter. This ID is obtained from the page snapshot.Returns:- Dict[str, str]: Execution result message.+ Dict[str, str]: A dictionary containing the result of the action.+ If pressing Enter causes a page navigation or DOM change, a+ "snapshot" key will be included with a diff of the changes."""self._validate_ref(ref, "enter")action: Dict[str, Any] = {"type": "enter", "ref": ref}return await self._exec_with_snapshot(action)- async def wait_user(self, *, seconds: float = 1.0) -> Dict[str, str]:- r"""Pause execution for a given amount of *real* time and then- return a *full* page snapshot.-- This is a convenience wrapper around the existing wait action for- scenarios where you encounter a CAPTCHA or need to pause for manual- user input, and want to retrieve the complete DOM snapshot afterward.+ async def wait_user(+ self,+ timeout_sec: Optional[float] = None,+ ) -> Dict[str, str]:+ r"""Pauses the agent's execution and waits for human intervention.+ This is useful for tasks that require manual steps, like solving a+ CAPTCHA. The agent will print a message and wait for the user to+ press the Enter key in the console.Args:- seconds (float): How long to sleep, expressed in seconds. Must- be a positive number. (default: :obj:`1.0`)+ timeout_sec (Optional[float]): The maximum time in seconds to wait+ for the user. If `None`, it will wait indefinitely. Defaults+ to `None`. (default: :obj:`None`)Returns:- Dict[str, str]: Keys ``result`` and ``snapshot``.+ Dict[str, str]: A dictionary containing a result message and a+ full snapshot of the current page after the user has acted.+ The keys are "result" and "snapshot"."""- if seconds is None or seconds <= 0:- logger.error("wait_time(): 'seconds' must be a positive number")- return {- "result": "wait_time(): 'seconds' must be a positive number"- }- # Reuse underlying ActionExecutor's ``wait`` implementation (expects- # ms)- timeout_ms = int(seconds * 1000)- action: Dict[str, Any] = {"type": "wait", "timeout": timeout_ms}+ import asyncio++ prompt = (+ "🕑 Agent is waiting for human input. "+ "Complete the required action in the browser, then press Enter "+ "to continue..."+ )- # Execute the sleep via ActionExecutor (no snapshot diff expected)- result = await self._exec(action)+ logger.info(f"\n{prompt}\n")++ async def _await_enter():+ await asyncio.to_thread(input, ">>> Press Enter to resume <<<\n")++ try:+ if timeout_sec is not None:+ await asyncio.wait_for(_await_enter(), timeout=timeout_sec)+ result_msg = "User resumed."+ else:+ await _await_enter()+ result_msg = "User resumed."+ except asyncio.TimeoutError:+ result_msg = f"Timeout {timeout_sec}s reached, auto-resumed."- # Always return a *full* snapshot after the pausesnapshot = await self._session.get_snapshot(- force_refresh=True, diff_only=False+ force_refresh=True,+ diff_only=False,)- return {"result": result, "snapshot": snapshot}++ return {"result": result_msg, "snapshot": snapshot}# Helper to run through ActionExecutorasync def _exec(self, action: Dict[str, Any]) -> str:diff --git a/examples/toolkits/browser_non_visual_toolkit_example.py b/examples/toolkits/browser_non_visual_toolkit_example.pyindex bdffe362c..2e4093b04 100644--- a/examples/toolkits/browser_non_visual_toolkit_example.py+++ b/examples/toolkits/browser_non_visual_toolkit_example.py@@ -41,14 +41,12 @@ model_backend = ModelFactory.create(web_toolkit = BrowserNonVisualToolkit(headless=False,- user_data_dir=USER_DATA_DIR,)agent = ChatAgent(- system_message="""- """,model=model_backend,tools=[*web_toolkit.get_tools()],+ max_iteration=10,)TASK_PROMPT = """Click "Sign in to Google Account" on Google.diff --git a/examples/toolkits/edgeone_pages_mcp_toolkit.py b/examples/toolkits/edgeone_pages_mcp_toolkit.pynew file mode 100644index 000000000..c1fb5dda3--- /dev/null+++ b/examples/toolkits/edgeone_pages_mcp_toolkit.py@@ -0,0 +1,56 @@+# ========= Copyright 2023-2024 @ CAMEL-AI.org. All Rights Reserved. =========+# Licensed under the Apache License, Version 2.0 (the "License");+# you may not use this file except in compliance with the License.+# You may obtain a copy of the License at+#+# http://www.apache.org/licenses/LICENSE-2.0+#+# Unless required by applicable law or agreed to in writing, software+# distributed under the License is distributed on an "AS IS" BASIS,+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.+# See the License for the specific language governing permissions and+# limitations under the License.+# ========= Copyright 2023-2024 @ CAMEL-AI.org. All Rights Reserved. =========+import asyncio++from camel.agents import ChatAgent+from camel.models import ModelFactory+from camel.toolkits import EdgeOnePagesMCPToolkit+from camel.types import ModelPlatformType, ModelType+++async def main():+ edgeone_pages_mcp_toolkit = EdgeOnePagesMCPToolkit()++ # connect to edgeone pages mcp+ await edgeone_pages_mcp_toolkit.connect()++ model = ModelFactory.create(+ model_platform=ModelPlatformType.DEFAULT,+ model_type=ModelType.DEFAULT,+ )++ chat_agent = ChatAgent(+ model=model,+ tools=edgeone_pages_mcp_toolkit.get_tools(),+ )++ response = await chat_agent.astep(+ "write a hello world page",+ )+ print(response.msg.content)++ # disconnect from edgeone pages mcp+ await edgeone_pages_mcp_toolkit.disconnect()+++# Run the async main function+if __name__ == "__main__":+ asyncio.run(main())++"""+==============================================================================+I have created a "Hello, World!" web page for you. You can view it here:+https://mcp.edgeone.site/share/7D8SNj2ftY9SRtPrWZ5rW+==============================================================================+"""diff --git a/examples/workforce/eigent.py b/examples/workforce/eigent.pyindex b7226fd49..b196bac65 100644--- a/examples/workforce/eigent.py+++ b/examples/workforce/eigent.py@@ -25,6 +25,7 @@ from camel.toolkits import (BrowserNonVisualToolkit,CodeExecutionToolkit,DalleToolkit,+ EdgeOnePagesMCPToolkit,FileWriteToolkit,FunctionTool,HumanToolkit,@@ -43,13 +44,17 @@ from camel.toolkits import (from camel.types import ModelPlatformType, ModelType-def developer_agent_factory(model: BaseModelBackend, task_id: str):+def developer_agent_factory(+ model: BaseModelBackend,+ task_id: str,+ edgeone_pages_mcp_toolkit: EdgeOnePagesMCPToolkit,+):r"""Factory for creating a developer agent."""tools = [- *TerminalToolkit().get_tools(),HumanToolkit().ask_human_via_console,*TerminalToolkit(clone_current_env=True).get_tools(),*CodeExecutionToolkit().get_tools(),+ *edgeone_pages_mcp_toolkit.get_tools(),]system_message = """You are a skilled coding assistant with DIRECT CODE@@ -62,6 +67,10 @@ def developer_agent_factory(model: BaseModelBackend, task_id: str):needed for efficient solutions- IMPLEMENT complete, production-ready code rather than theoreticalexamples+ - USE edgeone pages mcp toolkit to create and edit web pages, after you+ have created the web page, you can use the browser toolkit to view the+ web page,you can let search agent to visit the web page and verify the+ web page.- DEMONSTRATE results with proper error handling and practicalimplementation- If there's dependency issues when you try to execute code, you should@@ -339,94 +348,132 @@ When assisting users, always:async def main():- # Create a single model backend for all agents- model_backend = ModelFactory.create(- model_platform=ModelPlatformType.OPENAI,- model_type=ModelType.GPT_4_1_MINI,- # model_config_dict={- # "max_tokens": 32768,- # },- )-- model_backend_reason = ModelFactory.create(- model_platform=ModelPlatformType.OPENAI,- model_type=ModelType.GPT_4_1_MINI,- # model_config_dict={- # "max_tokens": 32768,- # },- )-- task_id = 'workforce_task'-- # Create agents using factory functions- search_agent = search_agent_factory(model_backend, task_id)- developer_agent = developer_agent_factory(model_backend, task_id)- document_agent = document_agent_factory(model_backend, task_id)- multi_modal_agent = multi_modal_agent_factory(model_backend, task_id)- # social_medium_agent = social_medium_agent_factory(model_backend,- # task_id)-- # Configure kwargs for all agents to use the same model_backend- coordinator_agent_kwargs = {"model": model_backend_reason}- task_agent_kwargs = {"model": model_backend_reason}- new_worker_agent_kwargs = {"model": model_backend}-- workforce = Workforce(- 'A workforce',- graceful_shutdown_timeout=30.0, # 30 seconds for debugging- share_memory=False,- coordinator_agent_kwargs=coordinator_agent_kwargs,- task_agent_kwargs=task_agent_kwargs,- new_worker_agent_kwargs=new_worker_agent_kwargs,- )-- workforce.add_single_agent_worker(- "Search Agent: Can search the web, extract webpage content, "- "simulate browser actions, and provide relevant information to "- "solve the given task.",- worker=search_agent,- ).add_single_agent_worker(- "Developer Agent: A skilled coding assistant that can write and "- "execute code, run terminal commands, and verify solutions to "- "complete tasks.",- worker=developer_agent,- ).add_single_agent_worker(- "Document Agent: A document processing assistant for creating, "- "modifying, and managing various document formats, including "- "presentations.",- worker=document_agent,- ).add_single_agent_worker(- "Multi-Modal Agent: A multi-modal processing assistant for "- "analyzing, and generating media content like audio and images.",- worker=multi_modal_agent,- )-- # specify the task to be solved- human_task = Task(- content=(- """-I want to read papers about GUI Agent. Please help me find ten papers and summarize them into a report.- """ # noqa: E501- ),- id='0',- )-- # Use the async version directly to avoid hanging with async tools- await workforce.process_task_async(human_task)-- # Test WorkforceLogger features- print("\n--- Workforce Log Tree ---")- print(workforce.get_workforce_log_tree())-- print("\n--- Workforce KPIs ---")- kpis = workforce.get_workforce_kpis()- for key, value in kpis.items():- print(f"{key}: {value}")-- log_file_path = "eigent_logs.json"- print(f"\n--- Dumping Workforce Logs to {log_file_path} ---")- workforce.dump_workforce_logs(log_file_path)- print(f"Logs dumped. Please check the file: {log_file_path}")+ edgeone_pages_mcp_toolkit = EdgeOnePagesMCPToolkit()+ try:+ await edgeone_pages_mcp_toolkit.connect()++ # Create a single model backend for all agents+ model_backend = ModelFactory.create(+ model_platform=ModelPlatformType.OPENAI,+ model_type=ModelType.GPT_4_1_MINI,+ # model_config_dict={+ # "max_tokens": 32768,+ # },+ )++ model_backend_reason = ModelFactory.create(+ model_platform=ModelPlatformType.OPENAI,+ model_type=ModelType.GPT_4_1_MINI,+ # model_config_dict={+ # "max_tokens": 32768,+ # },+ )++ task_id = 'workforce_task'++ # Configure kwargs for all agents to use the same model_backend+ coordinator_agent_kwargs = {"model": model_backend_reason}+ task_agent_kwargs = {"model": model_backend_reason}+ new_worker_agent_kwargs = {"model": model_backend}++ # Create agents using factory functions+ search_agent = search_agent_factory(+ model_backend, task_id, edgeone_pages_mcp_toolkit+ )+ developer_agent = developer_agent_factory(model_backend, task_id)+ document_agent = document_agent_factory(model_backend, task_id)+ multi_modal_agent = multi_modal_agent_factory(model_backend, task_id)++ # Configure kwargs for all agents to use the same model_backend+ coordinator_agent_kwargs = {"model": model_backend_reason}+ task_agent_kwargs = {"model": model_backend_reason}+ new_worker_agent_kwargs = {"model": model_backend}++ # Create workforce instance before adding workers+ workforce = Workforce(+ 'A workforce',+ graceful_shutdown_timeout=30.0, # 30 seconds for debugging+ share_memory=False,+ coordinator_agent_kwargs=coordinator_agent_kwargs,+ task_agent_kwargs=task_agent_kwargs,+ new_worker_agent_kwargs=new_worker_agent_kwargs,+ )++ workforce.add_single_agent_worker(+ "Search Agent: Can search the web, extract webpage content, "+ "simulate browser actions, and provide relevant information to "+ "solve the given task.",+ worker=search_agent,+ ).add_single_agent_worker(+ "Developer Agent: A skilled coding assistant that can write and "+ "execute code, run terminal commands, and verify solutions to "+ "complete tasks.",+ worker=developer_agent,+ ).add_single_agent_worker(+ "Document Agent: A document processing assistant for creating, "+ "modifying, and managing various document formats, including "+ "presentations.",+ worker=document_agent,+ ).add_single_agent_worker(+ "Multi-Modal Agent: A multi-modal processing assistant for "+ "analyzing, and generating media content like audio and images.",+ worker=multi_modal_agent,+ )++ # specify the task to be solved+ human_task = Task(+ content=(+ """+ I want to read papers about GUI Agent. Please help me find ten papers and summarize them into a report.+ """ # noqa: E501+ ),+ id='0',+ )++ workforce.add_single_agent_worker(+ "Search Agent: Can search the web, extract webpage content, "+ "simulate browser actions, and provide relevant information to "+ "solve the given task.",+ worker=search_agent,+ )++ workforce.add_single_agent_worker(+ "Document Agent: A document processing assistant for creating, "+ "modifying, and managing various document formats, including "+ "presentations.",+ worker=document_agent,+ )++ # specify the task to be solved+ human_task = Task(+ content=(+ """+ Gather today's headline news from BBC, Washington Post, Google News and+ CNN. Each headline can be processed in parallel by the SearchAgent pool.+ """+ ),+ id='0',+ )++ # Use the async version directly to avoid hanging with async tools+ await workforce.process_task_async(human_task)++ # Test WorkforceLogger features+ print("\n--- Workforce Log Tree ---")+ print(workforce.get_workforce_log_tree())++ print("\n--- Workforce KPIs ---")+ kpis = workforce.get_workforce_kpis()+ for key, value in kpis.items():+ print(f"{key}: {value}")++ log_file_path = "eigent_logs.json"+ print(f"\n--- Dumping Workforce Logs to {log_file_path} ---")+ workforce.dump_workforce_logs(log_file_path)+ print(f"Logs dumped. Please check the file: {log_file_path}")++ finally:+ await edgeone_pages_mcp_toolkit.disconnect()if __name__ == "__main__":