chore(boto3): restore boto3 dependency

Remove merge conflict hint code

.github/ISSUE_TEMPLATE/request_new_features.md

@@ -1,14 +0,0 @@
----
-name: "🤔 Request new features"
-about: Suggest ideas or features you’d like to see implemented in OpenManus.
-title: ''
-labels: kind/features
-assignees: ''
----
-
-**Feature description**
-<!-- Provide a clear and concise description of the proposed feature -->
-
-**Your Feature**
-<!-- Explain your idea or implementation process. Optionally, include a Pull Request URL. -->
-<!-- Ensure accompanying docs/tests/examples are provided for review. -->

.github/ISSUE_TEMPLATE/request_new_features.yaml

@@ -0,0 +1,21 @@
+name: "🤔 Request new features"
+description: Suggest ideas or features you’d like to see implemented in OpenManus.
+labels: enhancement
+body:
+  - type: textarea
+    id: feature-description
+    attributes:
+      label: Feature description
+      description: |
+        Provide a clear and concise description of the proposed feature
+    validations:
+      required: true
+  - type: textarea
+    id: your-feature
+    attributes:
+      label: Your Feature
+      description: |
+        Explain your idea or implementation process, if any. Optionally, include a Pull Request URL.
+        Ensure accompanying docs/tests/examples are provided for review.
+    validations:
+      required: false

.github/ISSUE_TEMPLATE/show_me_the_bug.md

@@ -1,25 +0,0 @@
----
-name: "🪲 Show me the Bug"
-about: Report a bug encountered while using OpenManus and seek assistance.
-title: ''
-labels: kind/bug
-assignees: ''
----
-
-**Bug description**
-<!-- Clearly describe the bug you encountered -->
-
-**Bug solved method**
-<!-- If resolved, explain the solution. Optionally, include a Pull Request URL. -->
-<!-- If unresolved, provide additional details to aid investigation -->
-
-**Environment information**
-<!-- System: e.g., Ubuntu 22.04, Python: e.g., 3.12, OpenManus version: e.g., 0.1.0 -->
-
-- System version:
-- Python version:
-- OpenManus version or branch:
-- Installation method (e.g., `pip install -r requirements.txt` or `pip install -e .`):
-
-**Screenshots or logs**
-<!-- Attach screenshots or logs to help diagnose the issue -->

.github/ISSUE_TEMPLATE/show_me_the_bug.yaml

@@ -0,0 +1,44 @@
+name: "🪲 Show me the Bug"
+description: Report a bug encountered while using OpenManus and seek assistance.
+labels: bug
+body:
+  - type: textarea
+    id: bug-description
+    attributes:
+      label: Bug Description
+      description: |
+        Clearly describe the bug you encountered
+    validations:
+      required: true
+  - type: textarea
+    id: solve-method
+    attributes:
+      label: Bug solved method
+      description: |
+        If resolved, explain the solution. Optionally, include a Pull Request URL.
+        If unresolved, provide additional details to aid investigation
+    validations:
+      required: true
+  - type: textarea
+    id: environment-information
+    attributes:
+      label: Environment information
+      description: |
+        System: e.g., Ubuntu 22.04
+        Python: e.g., 3.12
+        OpenManus version: e.g., 0.1.0
+      value: |
+        - System version:
+        - Python version:
+        - OpenManus version or branch:
+        - Installation method (e.g., `pip install -r requirements.txt` or `pip install -e .`):
+    validations:
+      required: true
+  - type: textarea
+    id: extra-information
+    attributes:
+      label: Extra information
+      description: |
+        For example, attach screenshots or logs to help diagnose the issue
+    validations:
+      required: false

.github/workflows/pr-autodiff.yaml

@@ -15,21 +15,20 @@ jobs:
       (github.event_name == 'pull_request') ||
       (github.event_name == 'issue_comment' &&
        contains(github.event.comment.body, '!pr-diff') &&
-       (github.event.comment.author_association == 'COLLABORATOR' || github.event.comment.author_association == 'MEMBER' || github.event.comment.author_association == 'OWNER') &&
+       (github.event.comment.author_association == 'CONTRIBUTOR' || github.event.comment.author_association == 'COLLABORATOR' || github.event.comment.author_association == 'MEMBER' || github.event.comment.author_association == 'OWNER') &&
        github.event.issue.pull_request)
     steps:
       - name: Get PR head SHA
         id: get-pr-sha
         run: |
-          if [ "${{ github.event_name }}" == "pull_request" ]; then
+          PR_URL="${{ github.event.issue.pull_request.url || github.event.pull_request.url }}"
-            echo "pr_sha=${{ github.event.pull_request.head.sha }}" >> $GITHUB_OUTPUT
+          # https://api.github.com/repos/OpenManus/pulls/1
-            echo "Retrieved PR head SHA: ${{ github.event.pull_request.head.sha }}"
+          RESPONSE=$(curl -s -H "Authorization: Bearer ${{ secrets.GITHUB_TOKEN }}" $PR_URL)
-          else
+          SHA=$(echo $RESPONSE | jq -r '.head.sha')
-            PR_URL="${{ github.event.issue.pull_request.url }}"
+          TARGET_BRANCH=$(echo $RESPONSE | jq -r '.base.ref')
-            SHA=$(curl -s -H "Authorization: Bearer ${{ secrets.GITHUB_TOKEN }}" $PR_URL | jq -r '.head.sha')
+          echo "pr_sha=$SHA" >> $GITHUB_OUTPUT
-            echo "pr_sha=$SHA" >> $GITHUB_OUTPUT
+          echo "target_branch=$TARGET_BRANCH" >> $GITHUB_OUTPUT
-            echo "Retrieved PR head SHA from API: $SHA"
+          echo "Retrieved PR head SHA from API: $SHA, target branch: $TARGET_BRANCH"
-          fi
       - name: Check out code
         uses: actions/checkout@v4
         with:
@@ -49,6 +48,7 @@ jobs:
           OPENAI_BASE_URL: ${{ secrets.OPENAI_BASE_URL }}
           GH_TOKEN: ${{ github.token }}
           PR_NUMBER: ${{ github.event.pull_request.number || github.event.issue.number }}
+          TARGET_BRANCH: ${{ steps.get-pr-sha.outputs.target_branch }}
         run: |-
           cat << 'EOF' > /tmp/_workflow_core.py
           import os
@@ -59,7 +59,7 @@ jobs:
 
           def get_diff():
               result = subprocess.run(
-                  ['git', 'diff', 'origin/main...HEAD'],
+                  ['git', 'diff', 'origin/' + os.getenv('TARGET_BRANCH') + '...HEAD'],
                   capture_output=True, text=True, check=True)
               return '\n'.join(
                   line for line in result.stdout.split('\n')
@@ -86,6 +86,17 @@ jobs:
 
               ### Spelling/Offensive Content Check
               - No spelling mistakes or offensive content found in the code or comments.
+
+              ## 中文（简体）
+              - 新增了 `ABC` 类
+              - `foo` 模块中的 `f()` 行为已修复
+
+              ### 评论高亮
+              - `config.toml` 需要正确配置才能确保新功能正常运行。
+
+              ### 内容检查
+              - 没有发现代码或注释中的拼写错误或不当措辞。
+
           3. Highlight non-English comments
           4. Check for spelling/offensive content'''
 

README.md

@@ -81,6 +81,11 @@ source .venv/bin/activate  # On Unix/macOS
 uv pip install -r requirements.txt
 ```
 
+### Browser Automation Tool (Optional)
+```bash
+playwright install
+```
+
 ## Configuration
 
 OpenManus requires configuration for the LLM APIs it uses. Follow these steps to set up your configuration:
@@ -119,7 +124,12 @@ python main.py
 
 Then input your idea via terminal!
 
-For unstable version, you also can run:
+For MCP tool version, you can run:
+```bash
+python run_mcp.py
+```
+
+For unstable multi-agent version, you also can run:
 
 ```bash
 python run_flow.py

README_ja.md

@@ -81,6 +81,11 @@ source .venv/bin/activate  # Unix/macOSの場合
 uv pip install -r requirements.txt
 ```
 
+### ブラウザ自動化ツール（オプション）
+```bash
+playwright install
+```
+
 ## 設定
 
 OpenManusを使用するには、LLM APIの設定が必要です。以下の手順に従って設定してください：
@@ -119,7 +124,12 @@ python main.py
 
 その後、ターミナルからプロンプトを入力してください！
 
-開発中バージョンを試すには、以下を実行します：
+MCP ツールバージョンを使用する場合は、以下を実行します：
+```bash
+python run_mcp.py
+```
+
+開発中のマルチエージェントバージョンを試すには、以下を実行します：
 
 ```bash
 python run_flow.py

README_ko.md

@@ -81,6 +81,11 @@ source .venv/bin/activate  # Unix/macOS의 경우
 uv pip install -r requirements.txt
 ```
 
+### 브라우저 자동화 도구 (선택사항)
+```bash
+playwright install
+```
+
 ## 설정 방법
 
 OpenManus를 사용하려면 사용하는 LLM API에 대한 설정이 필요합니다. 아래 단계를 따라 설정을 완료하세요:
@@ -119,7 +124,12 @@ python main.py
 
 이후 터미널에서 아이디어를 작성하세요!
 
-unstable 버전을 실행하려면 아래 명령어를 사용할 수도 있습니다:
+MCP 도구 버전을 사용하려면 다음을 실행하세요:
+```bash
+python run_mcp.py
+```
+
+불안정한 멀티 에이전트 버전을 실행하려면 다음을 실행할 수 있습니다:
 
 ```bash
 python run_flow.py

README_zh.md

@@ -82,6 +82,11 @@ source .venv/bin/activate  # Unix/macOS 系统
 uv pip install -r requirements.txt
 ```
 
+### 浏览器自动化工具（可选）
+```bash
+playwright install
+```
+
 ## 配置说明
 
 OpenManus 需要配置使用的 LLM API，请按以下步骤设置：
@@ -120,7 +125,12 @@ python main.py
 
 然后通过终端输入你的创意！
 
-如需体验不稳定的开发版本，可运行：
+如需使用 MCP 工具版本，可运行：
+```bash
+python run_mcp.py
+```
+
+如需体验不稳定的多智能体版本，可运行：
 
 ```bash
 python run_flow.py

app/agent/__init__.py

@@ -1,5 +1,6 @@
 from app.agent.base import BaseAgent
 from app.agent.browser import BrowserAgent
+from app.agent.mcp import MCPAgent
 from app.agent.planning import PlanningAgent
 from app.agent.react import ReActAgent
 from app.agent.swe import SWEAgent
@@ -13,4 +14,5 @@ __all__ = [
     "ReActAgent",
     "SWEAgent",
     "ToolCallAgent",
+    "MCPAgent",
 ]

app/agent/mcp.py

@@ -0,0 +1,185 @@
+from typing import Any, Dict, List, Optional, Tuple
+
+from pydantic import Field
+
+from app.agent.toolcall import ToolCallAgent
+from app.logger import logger
+from app.prompt.mcp import MULTIMEDIA_RESPONSE_PROMPT, NEXT_STEP_PROMPT, SYSTEM_PROMPT
+from app.schema import AgentState, Message
+from app.tool.base import ToolResult
+from app.tool.mcp import MCPClients
+
+
+class MCPAgent(ToolCallAgent):
+    """Agent for interacting with MCP (Model Context Protocol) servers.
+
+    This agent connects to an MCP server using either SSE or stdio transport
+    and makes the server's tools available through the agent's tool interface.
+    """
+
+    name: str = "mcp_agent"
+    description: str = "An agent that connects to an MCP server and uses its tools."
+
+    system_prompt: str = SYSTEM_PROMPT
+    next_step_prompt: str = NEXT_STEP_PROMPT
+
+    # Initialize MCP tool collection
+    mcp_clients: MCPClients = Field(default_factory=MCPClients)
+    available_tools: MCPClients = None  # Will be set in initialize()
+
+    max_steps: int = 20
+    connection_type: str = "stdio"  # "stdio" or "sse"
+
+    # Track tool schemas to detect changes
+    tool_schemas: Dict[str, Dict[str, Any]] = Field(default_factory=dict)
+    _refresh_tools_interval: int = 5  # Refresh tools every N steps
+
+    # Special tool names that should trigger termination
+    special_tool_names: List[str] = Field(default_factory=lambda: ["terminate"])
+
+    async def initialize(
+        self,
+        connection_type: Optional[str] = None,
+        server_url: Optional[str] = None,
+        command: Optional[str] = None,
+        args: Optional[List[str]] = None,
+    ) -> None:
+        """Initialize the MCP connection.
+
+        Args:
+            connection_type: Type of connection to use ("stdio" or "sse")
+            server_url: URL of the MCP server (for SSE connection)
+            command: Command to run (for stdio connection)
+            args: Arguments for the command (for stdio connection)
+        """
+        if connection_type:
+            self.connection_type = connection_type
+
+        # Connect to the MCP server based on connection type
+        if self.connection_type == "sse":
+            if not server_url:
+                raise ValueError("Server URL is required for SSE connection")
+            await self.mcp_clients.connect_sse(server_url=server_url)
+        elif self.connection_type == "stdio":
+            if not command:
+                raise ValueError("Command is required for stdio connection")
+            await self.mcp_clients.connect_stdio(command=command, args=args or [])
+        else:
+            raise ValueError(f"Unsupported connection type: {self.connection_type}")
+
+        # Set available_tools to our MCP instance
+        self.available_tools = self.mcp_clients
+
+        # Store initial tool schemas
+        await self._refresh_tools()
+
+        # Add system message about available tools
+        tool_names = list(self.mcp_clients.tool_map.keys())
+        tools_info = ", ".join(tool_names)
+
+        # Add system prompt and available tools information
+        self.memory.add_message(
+            Message.system_message(
+                f"{self.system_prompt}\n\nAvailable MCP tools: {tools_info}"
+            )
+        )
+
+    async def _refresh_tools(self) -> Tuple[List[str], List[str]]:
+        """Refresh the list of available tools from the MCP server.
+
+        Returns:
+            A tuple of (added_tools, removed_tools)
+        """
+        if not self.mcp_clients.session:
+            return [], []
+
+        # Get current tool schemas directly from the server
+        response = await self.mcp_clients.session.list_tools()
+        current_tools = {tool.name: tool.inputSchema for tool in response.tools}
+
+        # Determine added, removed, and changed tools
+        current_names = set(current_tools.keys())
+        previous_names = set(self.tool_schemas.keys())
+
+        added_tools = list(current_names - previous_names)
+        removed_tools = list(previous_names - current_names)
+
+        # Check for schema changes in existing tools
+        changed_tools = []
+        for name in current_names.intersection(previous_names):
+            if current_tools[name] != self.tool_schemas.get(name):
+                changed_tools.append(name)
+
+        # Update stored schemas
+        self.tool_schemas = current_tools
+
+        # Log and notify about changes
+        if added_tools:
+            logger.info(f"Added MCP tools: {added_tools}")
+            self.memory.add_message(
+                Message.system_message(f"New tools available: {', '.join(added_tools)}")
+            )
+        if removed_tools:
+            logger.info(f"Removed MCP tools: {removed_tools}")
+            self.memory.add_message(
+                Message.system_message(
+                    f"Tools no longer available: {', '.join(removed_tools)}"
+                )
+            )
+        if changed_tools:
+            logger.info(f"Changed MCP tools: {changed_tools}")
+
+        return added_tools, removed_tools
+
+    async def think(self) -> bool:
+        """Process current state and decide next action."""
+        # Check MCP session and tools availability
+        if not self.mcp_clients.session or not self.mcp_clients.tool_map:
+            logger.info("MCP service is no longer available, ending interaction")
+            self.state = AgentState.FINISHED
+            return False
+
+        # Refresh tools periodically
+        if self.current_step % self._refresh_tools_interval == 0:
+            await self._refresh_tools()
+            # All tools removed indicates shutdown
+            if not self.mcp_clients.tool_map:
+                logger.info("MCP service has shut down, ending interaction")
+                self.state = AgentState.FINISHED
+                return False
+
+        # Use the parent class's think method
+        return await super().think()
+
+    async def _handle_special_tool(self, name: str, result: Any, **kwargs) -> None:
+        """Handle special tool execution and state changes"""
+        # First process with parent handler
+        await super()._handle_special_tool(name, result, **kwargs)
+
+        # Handle multimedia responses
+        if isinstance(result, ToolResult) and result.base64_image:
+            self.memory.add_message(
+                Message.system_message(
+                    MULTIMEDIA_RESPONSE_PROMPT.format(tool_name=name)
+                )
+            )
+
+    def _should_finish_execution(self, name: str, **kwargs) -> bool:
+        """Determine if tool execution should finish the agent"""
+        # Terminate if the tool name is 'terminate'
+        return name.lower() == "terminate"
+
+    async def cleanup(self) -> None:
+        """Clean up MCP connection when done."""
+        if self.mcp_clients.session:
+            await self.mcp_clients.disconnect()
+            logger.info("MCP connection closed")
+
+    async def run(self, request: Optional[str] = None) -> str:
+        """Run the agent with cleanup when done."""
+        try:
+            result = await super().run(request)
+            return result
+        finally:
+            # Ensure cleanup happens even if there's an error
+            await self.cleanup()

app/agent/swe.py

@@ -29,7 +29,8 @@ class SWEAgent(ToolCallAgent):
     async def think(self) -> bool:
         """Process current state and decide next action"""
         # Update working directory
-        self.working_dir = await self.bash.execute("pwd")
+        result = await self.bash.execute("pwd")
+        self.working_dir = result.output
         self.next_step_prompt = self.next_step_prompt.format(
             current_dir=self.working_dir
         )

app/agent/toolcall.py

@@ -71,40 +71,42 @@ class ToolCallAgent(ReActAgent):
                 return False
             raise
 
-        self.tool_calls = response.tool_calls
+        self.tool_calls = tool_calls = (
+            response.tool_calls if response and response.tool_calls else []
+        )
+        content = response.content if response and response.content else ""
 
         # Log response info
-        logger.info(f"✨ {self.name}'s thoughts: {response.content}")
+        logger.info(f"✨ {self.name}'s thoughts: {content}")
         logger.info(
-            f"🛠️ {self.name} selected {len(response.tool_calls) if response.tool_calls else 0} tools to use"
+            f"🛠️ {self.name} selected {len(tool_calls) if tool_calls else 0} tools to use"
         )
-        if response.tool_calls:
+        if tool_calls:
             logger.info(
-                f"🧰 Tools being prepared: {[call.function.name for call in response.tool_calls]}"
+                f"🧰 Tools being prepared: {[call.function.name for call in tool_calls]}"
-            )
-            logger.info(
-                f"🔧 Tool arguments: {response.tool_calls[0].function.arguments}"
             )
+            logger.info(f"🔧 Tool arguments: {tool_calls[0].function.arguments}")
 
         try:
+            if response is None:
+                raise RuntimeError("No response received from the LLM")
+
             # Handle different tool_choices modes
             if self.tool_choices == ToolChoice.NONE:
-                if response.tool_calls:
+                if tool_calls:
                     logger.warning(
                         f"🤔 Hmm, {self.name} tried to use tools when they weren't available!"
                     )
-                if response.content:
+                if content:
-                    self.memory.add_message(Message.assistant_message(response.content))
+                    self.memory.add_message(Message.assistant_message(content))
                     return True
                 return False
 
             # Create and add assistant message
             assistant_msg = (
-                Message.from_tool_calls(
+                Message.from_tool_calls(content=content, tool_calls=self.tool_calls)
-                    content=response.content, tool_calls=self.tool_calls
-                )
                 if self.tool_calls
-                else Message.assistant_message(response.content)
+                else Message.assistant_message(content)
             )
             self.memory.add_message(assistant_msg)
 
@@ -113,7 +115,7 @@ class ToolCallAgent(ReActAgent):
 
             # For 'auto' mode, continue with content if no commands but content exists
             if self.tool_choices == ToolChoice.AUTO and not self.tool_calls:
-                return bool(response.content)
+                return bool(content)
 
             return bool(self.tool_calls)
         except Exception as e:
@@ -209,7 +211,7 @@ class ToolCallAgent(ReActAgent):
             return f"Error: {error_msg}"
         except Exception as e:
             error_msg = f"⚠️ Tool '{name}' encountered a problem: {str(e)}"
-            logger.error(error_msg)
+            logger.exception(error_msg)
             return f"Error: {error_msg}"
 
     async def _handle_special_tool(self, name: str, result: Any, **kwargs):

app/bedrock.py

@@ -0,0 +1,334 @@
+import json
+import sys
+import time
+import uuid
+from datetime import datetime
+from typing import Dict, List, Literal, Optional
+
+import boto3
+
+
+# Global variables to track the current tool use ID across function calls
+# Tmp solution
+CURRENT_TOOLUSE_ID = None
+
+
+# Class to handle OpenAI-style response formatting
+class OpenAIResponse:
+    def __init__(self, data):
+        # Recursively convert nested dicts and lists to OpenAIResponse objects
+        for key, value in data.items():
+            if isinstance(value, dict):
+                value = OpenAIResponse(value)
+            elif isinstance(value, list):
+                value = [
+                    OpenAIResponse(item) if isinstance(item, dict) else item
+                    for item in value
+                ]
+            setattr(self, key, value)
+
+    def model_dump(self, *args, **kwargs):
+        # Convert object to dict and add timestamp
+        data = self.__dict__
+        data["created_at"] = datetime.now().isoformat()
+        return data
+
+
+# Main client class for interacting with Amazon Bedrock
+class BedrockClient:
+    def __init__(self):
+        # Initialize Bedrock client, you need to configure AWS env first
+        try:
+            self.client = boto3.client("bedrock-runtime")
+            self.chat = Chat(self.client)
+        except Exception as e:
+            print(f"Error initializing Bedrock client: {e}")
+            sys.exit(1)
+
+
+# Chat interface class
+class Chat:
+    def __init__(self, client):
+        self.completions = ChatCompletions(client)
+
+
+# Core class handling chat completions functionality
+class ChatCompletions:
+    def __init__(self, client):
+        self.client = client
+
+    def _convert_openai_tools_to_bedrock_format(self, tools):
+        # Convert OpenAI function calling format to Bedrock tool format
+        bedrock_tools = []
+        for tool in tools:
+            if tool.get("type") == "function":
+                function = tool.get("function", {})
+                bedrock_tool = {
+                    "toolSpec": {
+                        "name": function.get("name", ""),
+                        "description": function.get("description", ""),
+                        "inputSchema": {
+                            "json": {
+                                "type": "object",
+                                "properties": function.get("parameters", {}).get(
+                                    "properties", {}
+                                ),
+                                "required": function.get("parameters", {}).get(
+                                    "required", []
+                                ),
+                            }
+                        },
+                    }
+                }
+                bedrock_tools.append(bedrock_tool)
+        return bedrock_tools
+
+    def _convert_openai_messages_to_bedrock_format(self, messages):
+        # Convert OpenAI message format to Bedrock message format
+        bedrock_messages = []
+        system_prompt = []
+        for message in messages:
+            if message.get("role") == "system":
+                system_prompt = [{"text": message.get("content")}]
+            elif message.get("role") == "user":
+                bedrock_message = {
+                    "role": message.get("role", "user"),
+                    "content": [{"text": message.get("content")}],
+                }
+                bedrock_messages.append(bedrock_message)
+            elif message.get("role") == "assistant":
+                bedrock_message = {
+                    "role": "assistant",
+                    "content": [{"text": message.get("content")}],
+                }
+                openai_tool_calls = message.get("tool_calls", [])
+                if openai_tool_calls:
+                    bedrock_tool_use = {
+                        "toolUseId": openai_tool_calls[0]["id"],
+                        "name": openai_tool_calls[0]["function"]["name"],
+                        "input": json.loads(
+                            openai_tool_calls[0]["function"]["arguments"]
+                        ),
+                    }
+                    bedrock_message["content"].append({"toolUse": bedrock_tool_use})
+                    global CURRENT_TOOLUSE_ID
+                    CURRENT_TOOLUSE_ID = openai_tool_calls[0]["id"]
+                bedrock_messages.append(bedrock_message)
+            elif message.get("role") == "tool":
+                bedrock_message = {
+                    "role": "user",
+                    "content": [
+                        {
+                            "toolResult": {
+                                "toolUseId": CURRENT_TOOLUSE_ID,
+                                "content": [{"text": message.get("content")}],
+                            }
+                        }
+                    ],
+                }
+                bedrock_messages.append(bedrock_message)
+            else:
+                raise ValueError(f"Invalid role: {message.get('role')}")
+        return system_prompt, bedrock_messages
+
+    def _convert_bedrock_response_to_openai_format(self, bedrock_response):
+        # Convert Bedrock response format to OpenAI format
+        content = ""
+        if bedrock_response.get("output", {}).get("message", {}).get("content"):
+            content_array = bedrock_response["output"]["message"]["content"]
+            content = "".join(item.get("text", "") for item in content_array)
+        if content == "":
+            content = "."
+
+        # Handle tool calls in response
+        openai_tool_calls = []
+        if bedrock_response.get("output", {}).get("message", {}).get("content"):
+            for content_item in bedrock_response["output"]["message"]["content"]:
+                if content_item.get("toolUse"):
+                    bedrock_tool_use = content_item["toolUse"]
+                    global CURRENT_TOOLUSE_ID
+                    CURRENT_TOOLUSE_ID = bedrock_tool_use["toolUseId"]
+                    openai_tool_call = {
+                        "id": CURRENT_TOOLUSE_ID,
+                        "type": "function",
+                        "function": {
+                            "name": bedrock_tool_use["name"],
+                            "arguments": json.dumps(bedrock_tool_use["input"]),
+                        },
+                    }
+                    openai_tool_calls.append(openai_tool_call)
+
+        # Construct final OpenAI format response
+        openai_format = {
+            "id": f"chatcmpl-{uuid.uuid4()}",
+            "created": int(time.time()),
+            "object": "chat.completion",
+            "system_fingerprint": None,
+            "choices": [
+                {
+                    "finish_reason": bedrock_response.get("stopReason", "end_turn"),
+                    "index": 0,
+                    "message": {
+                        "content": content,
+                        "role": bedrock_response.get("output", {})
+                        .get("message", {})
+                        .get("role", "assistant"),
+                        "tool_calls": openai_tool_calls
+                        if openai_tool_calls != []
+                        else None,
+                        "function_call": None,
+                    },
+                }
+            ],
+            "usage": {
+                "completion_tokens": bedrock_response.get("usage", {}).get(
+                    "outputTokens", 0
+                ),
+                "prompt_tokens": bedrock_response.get("usage", {}).get(
+                    "inputTokens", 0
+                ),
+                "total_tokens": bedrock_response.get("usage", {}).get("totalTokens", 0),
+            },
+        }
+        return OpenAIResponse(openai_format)
+
+    async def _invoke_bedrock(
+        self,
+        model: str,
+        messages: List[Dict[str, str]],
+        max_tokens: int,
+        temperature: float,
+        tools: Optional[List[dict]] = None,
+        tool_choice: Literal["none", "auto", "required"] = "auto",
+        **kwargs,
+    ) -> OpenAIResponse:
+        # Non-streaming invocation of Bedrock model
+        (
+            system_prompt,
+            bedrock_messages,
+        ) = self._convert_openai_messages_to_bedrock_format(messages)
+        response = self.client.converse(
+            modelId=model,
+            system=system_prompt,
+            messages=bedrock_messages,
+            inferenceConfig={"temperature": temperature, "maxTokens": max_tokens},
+            toolConfig={"tools": tools} if tools else None,
+        )
+        openai_response = self._convert_bedrock_response_to_openai_format(response)
+        return openai_response
+
+    async def _invoke_bedrock_stream(
+        self,
+        model: str,
+        messages: List[Dict[str, str]],
+        max_tokens: int,
+        temperature: float,
+        tools: Optional[List[dict]] = None,
+        tool_choice: Literal["none", "auto", "required"] = "auto",
+        **kwargs,
+    ) -> OpenAIResponse:
+        # Streaming invocation of Bedrock model
+        (
+            system_prompt,
+            bedrock_messages,
+        ) = self._convert_openai_messages_to_bedrock_format(messages)
+        response = self.client.converse_stream(
+            modelId=model,
+            system=system_prompt,
+            messages=bedrock_messages,
+            inferenceConfig={"temperature": temperature, "maxTokens": max_tokens},
+            toolConfig={"tools": tools} if tools else None,
+        )
+
+        # Initialize response structure
+        bedrock_response = {
+            "output": {"message": {"role": "", "content": []}},
+            "stopReason": "",
+            "usage": {},
+            "metrics": {},
+        }
+        bedrock_response_text = ""
+        bedrock_response_tool_input = ""
+
+        # Process streaming response
+        stream = response.get("stream")
+        if stream:
+            for event in stream:
+                if event.get("messageStart", {}).get("role"):
+                    bedrock_response["output"]["message"]["role"] = event[
+                        "messageStart"
+                    ]["role"]
+                if event.get("contentBlockDelta", {}).get("delta", {}).get("text"):
+                    bedrock_response_text += event["contentBlockDelta"]["delta"]["text"]
+                    print(
+                        event["contentBlockDelta"]["delta"]["text"], end="", flush=True
+                    )
+                if event.get("contentBlockStop", {}).get("contentBlockIndex") == 0:
+                    bedrock_response["output"]["message"]["content"].append(
+                        {"text": bedrock_response_text}
+                    )
+                if event.get("contentBlockStart", {}).get("start", {}).get("toolUse"):
+                    bedrock_tool_use = event["contentBlockStart"]["start"]["toolUse"]
+                    tool_use = {
+                        "toolUseId": bedrock_tool_use["toolUseId"],
+                        "name": bedrock_tool_use["name"],
+                    }
+                    bedrock_response["output"]["message"]["content"].append(
+                        {"toolUse": tool_use}
+                    )
+                    global CURRENT_TOOLUSE_ID
+                    CURRENT_TOOLUSE_ID = bedrock_tool_use["toolUseId"]
+                if event.get("contentBlockDelta", {}).get("delta", {}).get("toolUse"):
+                    bedrock_response_tool_input += event["contentBlockDelta"]["delta"][
+                        "toolUse"
+                    ]["input"]
+                    print(
+                        event["contentBlockDelta"]["delta"]["toolUse"]["input"],
+                        end="",
+                        flush=True,
+                    )
+                if event.get("contentBlockStop", {}).get("contentBlockIndex") == 1:
+                    bedrock_response["output"]["message"]["content"][1]["toolUse"][
+                        "input"
+                    ] = json.loads(bedrock_response_tool_input)
+        print()
+        openai_response = self._convert_bedrock_response_to_openai_format(
+            bedrock_response
+        )
+        return openai_response
+
+    def create(
+        self,
+        model: str,
+        messages: List[Dict[str, str]],
+        max_tokens: int,
+        temperature: float,
+        stream: Optional[bool] = True,
+        tools: Optional[List[dict]] = None,
+        tool_choice: Literal["none", "auto", "required"] = "auto",
+        **kwargs,
+    ) -> OpenAIResponse:
+        # Main entry point for chat completion
+        bedrock_tools = []
+        if tools is not None:
+            bedrock_tools = self._convert_openai_tools_to_bedrock_format(tools)
+        if stream:
+            return self._invoke_bedrock_stream(
+                model,
+                messages,
+                max_tokens,
+                temperature,
+                bedrock_tools,
+                tool_choice,
+                **kwargs,
+            )
+        else:
+            return self._invoke_bedrock(
+                model,
+                messages,
+                max_tokens,
+                temperature,
+                bedrock_tools,
+                tool_choice,
+                **kwargs,
+            )

app/config.py

@@ -25,7 +25,7 @@ class LLMSettings(BaseModel):
         description="Maximum input tokens to use across all requests (None for unlimited)",
     )
     temperature: float = Field(1.0, description="Sampling temperature")
-    api_type: str = Field(..., description="AzureOpenai or Openai")
+    api_type: str = Field(..., description="Azure, Openai, or Ollama")
     api_version: str = Field(..., description="Azure Openai version if AzureOpenai")
 
 
@@ -37,6 +37,18 @@ class ProxySettings(BaseModel):
 
 class SearchSettings(BaseModel):
     engine: str = Field(default="Google", description="Search engine the llm to use")
+    fallback_engines: List[str] = Field(
+        default_factory=lambda: ["DuckDuckGo", "Baidu"],
+        description="Fallback search engines to try if the primary engine fails",
+    )
+    retry_delay: int = Field(
+        default=60,
+        description="Seconds to wait before retrying all engines again after they all fail",
+    )
+    max_retries: int = Field(
+        default=3,
+        description="Maximum number of times to retry all engines when all fail",
+    )
 
 
 class BrowserSettings(BaseModel):
@@ -227,5 +239,10 @@ class Config:
         """Get the workspace root directory"""
         return WORKSPACE_ROOT
 
+    @property
+    def root_path(self) -> Path:
+        """Get the root path of the application"""
+        return PROJECT_ROOT
+
 
 config = Config()

app/flow/base.py

@@ -1,5 +1,4 @@
 from abc import ABC, abstractmethod
-from enum import Enum
 from typing import Dict, List, Optional, Union
 
 from pydantic import BaseModel
@@ -7,10 +6,6 @@ from pydantic import BaseModel
 from app.agent.base import BaseAgent
 
 
-class FlowType(str, Enum):
-    PLANNING = "planning"
-
-
 class BaseFlow(BaseModel, ABC):
     """Base class for execution flows supporting multiple agents"""
 
@@ -60,32 +55,3 @@ class BaseFlow(BaseModel, ABC):
     @abstractmethod
     async def execute(self, input_text: str) -> str:
         """Execute the flow with given input"""
-
-
-class PlanStepStatus(str, Enum):
-    """Enum class defining possible statuses of a plan step"""
-
-    NOT_STARTED = "not_started"
-    IN_PROGRESS = "in_progress"
-    COMPLETED = "completed"
-    BLOCKED = "blocked"
-
-    @classmethod
-    def get_all_statuses(cls) -> list[str]:
-        """Return a list of all possible step status values"""
-        return [status.value for status in cls]
-
-    @classmethod
-    def get_active_statuses(cls) -> list[str]:
-        """Return a list of values representing active statuses (not started or in progress)"""
-        return [cls.NOT_STARTED.value, cls.IN_PROGRESS.value]
-
-    @classmethod
-    def get_status_marks(cls) -> Dict[str, str]:
-        """Return a mapping of statuses to their marker symbols"""
-        return {
-            cls.COMPLETED.value: "[✓]",
-            cls.IN_PROGRESS.value: "[→]",
-            cls.BLOCKED.value: "[!]",
-            cls.NOT_STARTED.value: "[ ]",
-        }

app/flow/flow_factory.py

@@ -1,10 +1,15 @@
+from enum import Enum
 from typing import Dict, List, Union
 
 from app.agent.base import BaseAgent
-from app.flow.base import BaseFlow, FlowType
+from app.flow.base import BaseFlow
 from app.flow.planning import PlanningFlow
 
 
+class FlowType(str, Enum):
+    PLANNING = "planning"
+
+
 class FlowFactory:
     """Factory for creating different types of flows with support for multiple agents"""
 

app/flow/planning.py

@@ -1,17 +1,47 @@
 import json
 import time
+from enum import Enum
 from typing import Dict, List, Optional, Union
 
 from pydantic import Field
 
 from app.agent.base import BaseAgent
-from app.flow.base import BaseFlow, PlanStepStatus
+from app.flow.base import BaseFlow
 from app.llm import LLM
 from app.logger import logger
 from app.schema import AgentState, Message, ToolChoice
 from app.tool import PlanningTool
 
 
+class PlanStepStatus(str, Enum):
+    """Enum class defining possible statuses of a plan step"""
+
+    NOT_STARTED = "not_started"
+    IN_PROGRESS = "in_progress"
+    COMPLETED = "completed"
+    BLOCKED = "blocked"
+
+    @classmethod
+    def get_all_statuses(cls) -> list[str]:
+        """Return a list of all possible step status values"""
+        return [status.value for status in cls]
+
+    @classmethod
+    def get_active_statuses(cls) -> list[str]:
+        """Return a list of values representing active statuses (not started or in progress)"""
+        return [cls.NOT_STARTED.value, cls.IN_PROGRESS.value]
+
+    @classmethod
+    def get_status_marks(cls) -> Dict[str, str]:
+        """Return a mapping of statuses to their marker symbols"""
+        return {
+            cls.COMPLETED.value: "[✓]",
+            cls.IN_PROGRESS.value: "[→]",
+            cls.BLOCKED.value: "[!]",
+            cls.NOT_STARTED.value: "[ ]",
+        }
+
+
 class PlanningFlow(BaseFlow):
     """A flow that manages planning and execution of tasks using agents."""
 

app/llm.py

@@ -10,6 +10,7 @@ from openai import (
     OpenAIError,
     RateLimitError,
 )
+from openai.types.chat.chat_completion_message import ChatCompletionMessage
 from tenacity import (
     retry,
     retry_if_exception_type,
@@ -17,6 +18,7 @@ from tenacity import (
     wait_random_exponential,
 )
 
+from app.bedrock import BedrockClient
 from app.config import LLMSettings, config
 from app.exceptions import TokenLimitExceeded
 from app.logger import logger  # Assuming a logger is set up in your app
@@ -30,6 +32,14 @@ from app.schema import (
 
 
 REASONING_MODELS = ["o1", "o3-mini"]
+MULTIMODAL_MODELS = [
+    "gpt-4-vision-preview",
+    "gpt-4o",
+    "gpt-4o-mini",
+    "claude-3-opus-20240229",
+    "claude-3-sonnet-20240229",
+    "claude-3-haiku-20240307",
+]
 
 
 class TokenCounter:
@@ -216,6 +226,8 @@ class LLM:
                     api_key=self.api_key,
                     api_version=self.api_version,
                 )
+            elif self.api_type == "aws":
+                self.client = BedrockClient()
             else:
                 self.client = AsyncOpenAI(api_key=self.api_key, base_url=self.base_url)
 
@@ -259,12 +271,15 @@ class LLM:
         return "Token limit exceeded"
 
     @staticmethod
-    def format_messages(messages: List[Union[dict, Message]]) -> List[dict]:
+    def format_messages(
+        messages: List[Union[dict, Message]], supports_images: bool = False
+    ) -> List[dict]:
         """
         Format messages for LLM by converting them to OpenAI message format.
 
         Args:
             messages: List of messages that can be either dict or Message objects
+            supports_images: Flag indicating if the target model supports image inputs
 
         Returns:
             List[dict]: List of formatted messages in OpenAI format
@@ -288,54 +303,58 @@ class LLM:
             if isinstance(message, Message):
                 message = message.to_dict()
 
-            if not isinstance(message, dict):
+            if isinstance(message, dict):
+                # If message is a dict, ensure it has required fields
+                if "role" not in message:
+                    raise ValueError("Message dict must contain 'role' field")
+
+                # Process base64 images if present and model supports images
+                if supports_images and message.get("base64_image"):
+                    # Initialize or convert content to appropriate format
+                    if not message.get("content"):
+                        message["content"] = []
+                    elif isinstance(message["content"], str):
+                        message["content"] = [
+                            {"type": "text", "text": message["content"]}
+                        ]
+                    elif isinstance(message["content"], list):
+                        # Convert string items to proper text objects
+                        message["content"] = [
+                            (
+                                {"type": "text", "text": item}
+                                if isinstance(item, str)
+                                else item
+                            )
+                            for item in message["content"]
+                        ]
+
+                    # Add the image to content
+                    message["content"].append(
+                        {
+                            "type": "image_url",
+                            "image_url": {
+                                "url": f"data:image/jpeg;base64,{message['base64_image']}"
+                            },
+                        }
+                    )
+
+                    # Remove the base64_image field
+                    del message["base64_image"]
+                # If model doesn't support images but message has base64_image, handle gracefully
+                elif not supports_images and message.get("base64_image"):
+                    # Just remove the base64_image field and keep the text content
+                    del message["base64_image"]
+
+                if "content" in message or "tool_calls" in message:
+                    formatted_messages.append(message)
+                # else: do not include the message
+            else:
                 raise TypeError(f"Unsupported message type: {type(message)}")
 
-            # Validate required fields
+        # Validate all messages have required fields
-            if "role" not in message:
+        for msg in formatted_messages:
-                raise ValueError("Message dict must contain 'role' field")
+            if msg["role"] not in ROLE_VALUES:
-
+                raise ValueError(f"Invalid role: {msg['role']}")
-            # Process base64 images if present
-            if message.get("base64_image"):
-                # Initialize or convert content to appropriate format
-                if not message.get("content"):
-                    message["content"] = []
-                elif isinstance(message["content"], str):
-                    message["content"] = [{"type": "text", "text": message["content"]}]
-                elif isinstance(message["content"], list):
-                    # Convert string items to proper text objects
-                    message["content"] = [
-                        (
-                            {"type": "text", "text": item}
-                            if isinstance(item, str)
-                            else item
-                        )
-                        for item in message["content"]
-                    ]
-
-                # Add the image to content
-                message["content"].append(
-                    {
-                        "type": "image_url",
-                        "image_url": {
-                            "url": f"data:image/jpeg;base64,{message['base64_image']}"
-                        },
-                    }
-                )
-
-                # Remove the base64_image field
-                del message["base64_image"]
-
-            # Only include messages with content or tool_calls
-            if "content" in message or "tool_calls" in message:
-                formatted_messages.append(message)
-
-        # Validate all roles
-        invalid_roles = [
-            msg for msg in formatted_messages if msg["role"] not in ROLE_VALUES
-        ]
-        if invalid_roles:
-            raise ValueError(f"Invalid role: {invalid_roles[0]['role']}")
 
         return formatted_messages
 
@@ -372,12 +391,15 @@ class LLM:
             Exception: For unexpected errors
         """
         try:
-            # Format system and user messages
+            # Check if the model supports images
+            supports_images = self.model in MULTIMODAL_MODELS
+
+            # Format system and user messages with image support check
             if system_msgs:
-                system_msgs = self.format_messages(system_msgs)
+                system_msgs = self.format_messages(system_msgs, supports_images)
-                messages = system_msgs + self.format_messages(messages)
+                messages = system_msgs + self.format_messages(messages, supports_images)
             else:
-                messages = self.format_messages(messages)
+                messages = self.format_messages(messages, supports_images)
 
             # Calculate input token count
             input_tokens = self.count_message_tokens(messages)
@@ -403,9 +425,9 @@ class LLM:
 
             if not stream:
                 # Non-streaming request
-                params["stream"] = False
+                response = await self.client.chat.completions.create(
-
+                    **params, stream=False
-                response = await self.client.chat.completions.create(**params)
+                )
 
                 if not response.choices or not response.choices[0].message.content:
                     raise ValueError("Empty or invalid response from LLM")
@@ -420,8 +442,7 @@ class LLM:
             # Streaming request, For streaming, update estimated token count before making the request
             self.update_token_count(input_tokens)
 
-            params["stream"] = True
+            response = await self.client.chat.completions.create(**params, stream=True)
-            response = await self.client.chat.completions.create(**params)
 
             collected_messages = []
             completion_text = ""
@@ -448,11 +469,11 @@ class LLM:
         except TokenLimitExceeded:
             # Re-raise token limit errors without logging
             raise
-        except ValueError as ve:
+        except ValueError:
-            logger.error(f"Validation error: {ve}")
+            logger.exception(f"Validation error")
             raise
         except OpenAIError as oe:
-            logger.error(f"OpenAI API error: {oe}")
+            logger.exception(f"OpenAI API error")
             if isinstance(oe, AuthenticationError):
                 logger.error("Authentication failed. Check API key.")
             elif isinstance(oe, RateLimitError):
@@ -460,8 +481,8 @@ class LLM:
             elif isinstance(oe, APIError):
                 logger.error(f"API error: {oe}")
             raise
-        except Exception as e:
+        except Exception:
-            logger.error(f"Unexpected error in ask: {e}")
+            logger.exception(f"Unexpected error in ask")
             raise
 
     @retry(
@@ -499,8 +520,15 @@ class LLM:
             Exception: For unexpected errors
         """
         try:
-            # Format messages
+            # For ask_with_images, we always set supports_images to True because
-            formatted_messages = self.format_messages(messages)
+            # this method should only be called with models that support images
+            if self.model not in MULTIMODAL_MODELS:
+                raise ValueError(
+                    f"Model {self.model} does not support images. Use a model from {MULTIMODAL_MODELS}"
+                )
+
+            # Format messages with image support
+            formatted_messages = self.format_messages(messages, supports_images=True)
 
             # Ensure the last message is from the user to attach images
             if not formatted_messages or formatted_messages[-1]["role"] != "user":
@@ -539,7 +567,10 @@ class LLM:
 
             # Add system messages if provided
             if system_msgs:
-                all_messages = self.format_messages(system_msgs) + formatted_messages
+                all_messages = (
+                    self.format_messages(system_msgs, supports_images=True)
+                    + formatted_messages
+                )
             else:
                 all_messages = formatted_messages
 
@@ -626,7 +657,7 @@ class LLM:
         tool_choice: TOOL_CHOICE_TYPE = ToolChoice.AUTO,  # type: ignore
         temperature: Optional[float] = None,
         **kwargs,
-    ):
+    ) -> ChatCompletionMessage | None:
         """
         Ask LLM using functions/tools and return the response.
 
@@ -653,12 +684,15 @@ class LLM:
             if tool_choice not in TOOL_CHOICE_VALUES:
                 raise ValueError(f"Invalid tool_choice: {tool_choice}")
 
+            # Check if the model supports images
+            supports_images = self.model in MULTIMODAL_MODELS
+
             # Format messages
             if system_msgs:
-                system_msgs = self.format_messages(system_msgs)
+                system_msgs = self.format_messages(system_msgs, supports_images)
-                messages = system_msgs + self.format_messages(messages)
+                messages = system_msgs + self.format_messages(messages, supports_images)
             else:
-                messages = self.format_messages(messages)
+                messages = self.format_messages(messages, supports_images)
 
             # Calculate input token count
             input_tokens = self.count_message_tokens(messages)
@@ -701,12 +735,15 @@ class LLM:
                     temperature if temperature is not None else self.temperature
                 )
 
-            response = await self.client.chat.completions.create(**params)
+            response: ChatCompletion = await self.client.chat.completions.create(
+                **params, stream=False
+            )
 
             # Check if response is valid
             if not response.choices or not response.choices[0].message:
                 print(response)
-                raise ValueError("Invalid or empty response from LLM")
+                # raise ValueError("Invalid or empty response from LLM")
+                return None
 
             # Update token counts
             self.update_token_count(

app/mcp/server.py

@@ -0,0 +1,180 @@
+import logging
+import sys
+
+
+logging.basicConfig(level=logging.INFO, handlers=[logging.StreamHandler(sys.stderr)])
+
+import argparse
+import asyncio
+import atexit
+import json
+from inspect import Parameter, Signature
+from typing import Any, Dict, Optional
+
+from mcp.server.fastmcp import FastMCP
+
+from app.logger import logger
+from app.tool.base import BaseTool
+from app.tool.bash import Bash
+from app.tool.browser_use_tool import BrowserUseTool
+from app.tool.str_replace_editor import StrReplaceEditor
+from app.tool.terminate import Terminate
+
+
+class MCPServer:
+    """MCP Server implementation with tool registration and management."""
+
+    def __init__(self, name: str = "openmanus"):
+        self.server = FastMCP(name)
+        self.tools: Dict[str, BaseTool] = {}
+
+        # Initialize standard tools
+        self.tools["bash"] = Bash()
+        self.tools["browser"] = BrowserUseTool()
+        self.tools["editor"] = StrReplaceEditor()
+        self.tools["terminate"] = Terminate()
+
+    def register_tool(self, tool: BaseTool, method_name: Optional[str] = None) -> None:
+        """Register a tool with parameter validation and documentation."""
+        tool_name = method_name or tool.name
+        tool_param = tool.to_param()
+        tool_function = tool_param["function"]
+
+        # Define the async function to be registered
+        async def tool_method(**kwargs):
+            logger.info(f"Executing {tool_name}: {kwargs}")
+            result = await tool.execute(**kwargs)
+
+            logger.info(f"Result of {tool_name}: {result}")
+
+            # Handle different types of results (match original logic)
+            if hasattr(result, "model_dump"):
+                return json.dumps(result.model_dump())
+            elif isinstance(result, dict):
+                return json.dumps(result)
+            return result
+
+        # Set method metadata
+        tool_method.__name__ = tool_name
+        tool_method.__doc__ = self._build_docstring(tool_function)
+        tool_method.__signature__ = self._build_signature(tool_function)
+
+        # Store parameter schema (important for tools that access it programmatically)
+        param_props = tool_function.get("parameters", {}).get("properties", {})
+        required_params = tool_function.get("parameters", {}).get("required", [])
+        tool_method._parameter_schema = {
+            param_name: {
+                "description": param_details.get("description", ""),
+                "type": param_details.get("type", "any"),
+                "required": param_name in required_params,
+            }
+            for param_name, param_details in param_props.items()
+        }
+
+        # Register with server
+        self.server.tool()(tool_method)
+        logger.info(f"Registered tool: {tool_name}")
+
+    def _build_docstring(self, tool_function: dict) -> str:
+        """Build a formatted docstring from tool function metadata."""
+        description = tool_function.get("description", "")
+        param_props = tool_function.get("parameters", {}).get("properties", {})
+        required_params = tool_function.get("parameters", {}).get("required", [])
+
+        # Build docstring (match original format)
+        docstring = description
+        if param_props:
+            docstring += "\n\nParameters:\n"
+            for param_name, param_details in param_props.items():
+                required_str = (
+                    "(required)" if param_name in required_params else "(optional)"
+                )
+                param_type = param_details.get("type", "any")
+                param_desc = param_details.get("description", "")
+                docstring += (
+                    f"    {param_name} ({param_type}) {required_str}: {param_desc}\n"
+                )
+
+        return docstring
+
+    def _build_signature(self, tool_function: dict) -> Signature:
+        """Build a function signature from tool function metadata."""
+        param_props = tool_function.get("parameters", {}).get("properties", {})
+        required_params = tool_function.get("parameters", {}).get("required", [])
+
+        parameters = []
+
+        # Follow original type mapping
+        for param_name, param_details in param_props.items():
+            param_type = param_details.get("type", "")
+            default = Parameter.empty if param_name in required_params else None
+
+            # Map JSON Schema types to Python types (same as original)
+            annotation = Any
+            if param_type == "string":
+                annotation = str
+            elif param_type == "integer":
+                annotation = int
+            elif param_type == "number":
+                annotation = float
+            elif param_type == "boolean":
+                annotation = bool
+            elif param_type == "object":
+                annotation = dict
+            elif param_type == "array":
+                annotation = list
+
+            # Create parameter with same structure as original
+            param = Parameter(
+                name=param_name,
+                kind=Parameter.KEYWORD_ONLY,
+                default=default,
+                annotation=annotation,
+            )
+            parameters.append(param)
+
+        return Signature(parameters=parameters)
+
+    async def cleanup(self) -> None:
+        """Clean up server resources."""
+        logger.info("Cleaning up resources")
+        # Follow original cleanup logic - only clean browser tool
+        if "browser" in self.tools and hasattr(self.tools["browser"], "cleanup"):
+            await self.tools["browser"].cleanup()
+
+    def register_all_tools(self) -> None:
+        """Register all tools with the server."""
+        for tool in self.tools.values():
+            self.register_tool(tool)
+
+    def run(self, transport: str = "stdio") -> None:
+        """Run the MCP server."""
+        # Register all tools
+        self.register_all_tools()
+
+        # Register cleanup function (match original behavior)
+        atexit.register(lambda: asyncio.run(self.cleanup()))
+
+        # Start server (with same logging as original)
+        logger.info(f"Starting OpenManus server ({transport} mode)")
+        self.server.run(transport=transport)
+
+
+def parse_args() -> argparse.Namespace:
+    """Parse command line arguments."""
+    parser = argparse.ArgumentParser(description="OpenManus MCP Server")
+    parser.add_argument(
+        "--transport",
+        choices=["stdio"],
+        default="stdio",
+        help="Communication method: stdio or http (default: stdio)",
+    )
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    args = parse_args()
+
+    # Create and run server (maintaining original flow)
+    server = MCPServer()
+    server.run(transport=args.transport)

app/prompt/mcp.py

@@ -0,0 +1,43 @@
+"""Prompts for the MCP Agent."""
+
+SYSTEM_PROMPT = """You are an AI assistant with access to a Model Context Protocol (MCP) server.
+You can use the tools provided by the MCP server to complete tasks.
+The MCP server will dynamically expose tools that you can use - always check the available tools first.
+
+When using an MCP tool:
+1. Choose the appropriate tool based on your task requirements
+2. Provide properly formatted arguments as required by the tool
+3. Observe the results and use them to determine next steps
+4. Tools may change during operation - new tools might appear or existing ones might disappear
+
+Follow these guidelines:
+- Call tools with valid parameters as documented in their schemas
+- Handle errors gracefully by understanding what went wrong and trying again with corrected parameters
+- For multimedia responses (like images), you'll receive a description of the content
+- Complete user requests step by step, using the most appropriate tools
+- If multiple tools need to be called in sequence, make one call at a time and wait for results
+
+Remember to clearly explain your reasoning and actions to the user.
+"""
+
+NEXT_STEP_PROMPT = """Based on the current state and available tools, what should be done next?
+Think step by step about the problem and identify which MCP tool would be most helpful for the current stage.
+If you've already made progress, consider what additional information you need or what actions would move you closer to completing the task.
+"""
+
+# Additional specialized prompts
+TOOL_ERROR_PROMPT = """You encountered an error with the tool '{tool_name}'.
+Try to understand what went wrong and correct your approach.
+Common issues include:
+- Missing or incorrect parameters
+- Invalid parameter formats
+- Using a tool that's no longer available
+- Attempting an operation that's not supported
+
+Please check the tool specifications and try again with corrected parameters.
+"""
+
+MULTIMEDIA_RESPONSE_PROMPT = """You've received a multimedia response (image, audio, etc.) from the tool '{tool_name}'.
+This content has been processed and described for you.
+Use this information to continue the task or provide insights to the user.
+"""

app/tool/bash.py

@@ -3,7 +3,7 @@ import os
 from typing import Optional
 
 from app.exceptions import ToolError
-from app.tool.base import BaseTool, CLIResult, ToolResult
+from app.tool.base import BaseTool, CLIResult
 
 
 _BASH_DESCRIPTION = """Execute a bash command in the terminal.
@@ -57,7 +57,7 @@ class _BashSession:
         if not self._started:
             raise ToolError("Session has not started.")
         if self._process.returncode is not None:
-            return ToolResult(
+            return CLIResult(
                 system="tool must be restarted",
                 error=f"bash has exited with returncode {self._process.returncode}",
             )
@@ -140,7 +140,7 @@ class Bash(BaseTool):
             self._session = _BashSession()
             await self._session.start()
 
-            return ToolResult(system="tool has been restarted.")
+            return CLIResult(system="tool has been restarted.")
 
         if self._session is None:
             self._session = _BashSession()

app/tool/browser_use_tool.py

@@ -1,4 +1,5 @@
 import asyncio
+import base64
 import json
 from typing import Generic, Optional, TypeVar
 
@@ -418,17 +419,7 @@ class BrowserUseTool(BaseTool, Generic[Context]):
 
                         # Create prompt for LLM
                         prompt_text = """
-Your task is to extract the content of the page. You will be given a page and a goal, and you should extract all relevant information around this goal from the page.
+Your task is to extract the content of the page. You will be given a page and a goal, and you should extract all relevant information around this goal from the page. If the goal is vague, summarize the page. Respond in json format.
-
-Examples of extraction goals:
-- Extract all company names
-- Extract specific descriptions
-- Extract all information about a topic
-- Extract links with companies in structured format
-- Extract all links
-
-If the goal is vague, summarize the page. Respond in JSON format.
-
 Extraction goal: {goal}
 
 Page content:
@@ -445,10 +436,54 @@ Page content:
 
                         messages = [Message.user_message(formatted_prompt)]
 
-                        # Use LLM to extract content based on the goal
+                        # Define extraction function for the tool
-                        response = await self.llm.ask(messages)
+                        extraction_function = {
+                            "type": "function",
+                            "function": {
+                                "name": "extract_content",
+                                "description": "Extract specific information from a webpage based on a goal",
+                                "parameters": {
+                                    "type": "object",
+                                    "properties": {
+                                        "extracted_content": {
+                                            "type": "object",
+                                            "description": "The content extracted from the page according to the goal",
+                                        }
+                                    },
+                                    "required": ["extracted_content"],
+                                },
+                            },
+                        }
+
+                        # Use LLM to extract content with required function calling
+                        response = await self.llm.ask_tool(
+                            messages,
+                            tools=[extraction_function],
+                            tool_choice="required",
+                        )
+
+                        # Extract content from function call response
+                        if (
+                            response
+                            and response.tool_calls
+                            and len(response.tool_calls) > 0
+                        ):
+                            # Get the first tool call arguments
+                            tool_call = response.tool_calls[0]
+                            # Parse the JSON arguments
+                            try:
+                                args = json.loads(tool_call.function.arguments)
+                                extracted_content = args.get("extracted_content", {})
+                                # Format extracted content as JSON string
+                                content_json = json.dumps(
+                                    extracted_content, indent=2, ensure_ascii=False
+                                )
+                                msg = f"Extracted from page:\n{content_json}\n"
+                            except Exception as e:
+                                msg = f"Error parsing extraction result: {str(e)}\nRaw response: {tool_call.function.arguments}"
+                        else:
+                            msg = "No content was extracted from the page."
 
-                        msg = f"Extracted from page:\n{response}\n"
                         return ToolResult(output=msg)
                     except Exception as e:
                         # Provide a more helpful error message
@@ -518,7 +553,16 @@ Page content:
                 viewport_height = ctx.config.browser_window_size.get("height", 0)
 
             # Take a screenshot for the state
-            screenshot = await ctx.take_screenshot(full_page=True)
+            page = await ctx.get_current_page()
+
+            await page.bring_to_front()
+            await page.wait_for_load_state()
+
+            screenshot = await page.screenshot(
+                full_page=True, animations="disabled", type="jpeg", quality=100
+            )
+
+            screenshot = base64.b64encode(screenshot).decode("utf-8")
 
             # Build the state info with all required fields
             state_info = {

app/tool/file_operators.py

@@ -42,17 +42,19 @@ class FileOperator(Protocol):
 class LocalFileOperator(FileOperator):
     """File operations implementation for local filesystem."""
 
+    encoding: str = "utf-8"
+
     async def read_file(self, path: PathLike) -> str:
         """Read content from a local file."""
         try:
-            return Path(path).read_text()
+            return Path(path).read_text(encoding=self.encoding)
         except Exception as e:
             raise ToolError(f"Failed to read {path}: {str(e)}") from None
 
     async def write_file(self, path: PathLike, content: str) -> None:
         """Write content to a local file."""
         try:
-            Path(path).write_text(content)
+            Path(path).write_text(content, encoding=self.encoding)
         except Exception as e:
             raise ToolError(f"Failed to write to {path}: {str(e)}") from None
 

app/tool/mcp.py

@@ -0,0 +1,115 @@
+from contextlib import AsyncExitStack
+from typing import List, Optional
+
+from mcp import ClientSession, StdioServerParameters
+from mcp.client.sse import sse_client
+from mcp.client.stdio import stdio_client
+from mcp.types import TextContent
+
+from app.logger import logger
+from app.tool.base import BaseTool, ToolResult
+from app.tool.tool_collection import ToolCollection
+
+
+class MCPClientTool(BaseTool):
+    """Represents a tool proxy that can be called on the MCP server from the client side."""
+
+    session: Optional[ClientSession] = None
+
+    async def execute(self, **kwargs) -> ToolResult:
+        """Execute the tool by making a remote call to the MCP server."""
+        if not self.session:
+            return ToolResult(error="Not connected to MCP server")
+
+        try:
+            result = await self.session.call_tool(self.name, kwargs)
+            content_str = ", ".join(
+                item.text for item in result.content if isinstance(item, TextContent)
+            )
+            return ToolResult(output=content_str or "No output returned.")
+        except Exception as e:
+            return ToolResult(error=f"Error executing tool: {str(e)}")
+
+
+class MCPClients(ToolCollection):
+    """
+    A collection of tools that connects to an MCP server and manages available tools through the Model Context Protocol.
+    """
+
+    session: Optional[ClientSession] = None
+    exit_stack: AsyncExitStack = None
+    description: str = "MCP client tools for server interaction"
+
+    def __init__(self):
+        super().__init__()  # Initialize with empty tools list
+        self.name = "mcp"  # Keep name for backward compatibility
+        self.exit_stack = AsyncExitStack()
+
+    async def connect_sse(self, server_url: str) -> None:
+        """Connect to an MCP server using SSE transport."""
+        if not server_url:
+            raise ValueError("Server URL is required.")
+        if self.session:
+            await self.disconnect()
+
+        streams_context = sse_client(url=server_url)
+        streams = await self.exit_stack.enter_async_context(streams_context)
+        self.session = await self.exit_stack.enter_async_context(
+            ClientSession(*streams)
+        )
+
+        await self._initialize_and_list_tools()
+
+    async def connect_stdio(self, command: str, args: List[str]) -> None:
+        """Connect to an MCP server using stdio transport."""
+        if not command:
+            raise ValueError("Server command is required.")
+        if self.session:
+            await self.disconnect()
+
+        server_params = StdioServerParameters(command=command, args=args)
+        stdio_transport = await self.exit_stack.enter_async_context(
+            stdio_client(server_params)
+        )
+        read, write = stdio_transport
+        self.session = await self.exit_stack.enter_async_context(
+            ClientSession(read, write)
+        )
+
+        await self._initialize_and_list_tools()
+
+    async def _initialize_and_list_tools(self) -> None:
+        """Initialize session and populate tool map."""
+        if not self.session:
+            raise RuntimeError("Session not initialized.")
+
+        await self.session.initialize()
+        response = await self.session.list_tools()
+
+        # Clear existing tools
+        self.tools = tuple()
+        self.tool_map = {}
+
+        # Create proper tool objects for each server tool
+        for tool in response.tools:
+            server_tool = MCPClientTool(
+                name=tool.name,
+                description=tool.description,
+                parameters=tool.inputSchema,
+                session=self.session,
+            )
+            self.tool_map[tool.name] = server_tool
+
+        self.tools = tuple(self.tool_map.values())
+        logger.info(
+            f"Connected to server with tools: {[tool.name for tool in response.tools]}"
+        )
+
+    async def disconnect(self) -> None:
+        """Disconnect from the MCP server and clean up resources."""
+        if self.session and self.exit_stack:
+            await self.exit_stack.aclose()
+            self.session = None
+            self.tools = tuple()
+            self.tool_map = {}
+            logger.info("Disconnected from MCP server")

app/tool/search/__init__.py

@@ -4,6 +4,7 @@ from app.tool.search.bing_search import BingSearchEngine
 from app.tool.search.duckduckgo_search import DuckDuckGoSearchEngine
 from app.tool.search.google_search import GoogleSearchEngine
 
+
 __all__ = [
     "WebSearchEngine",
     "BaiduSearchEngine",

app/tool/search/bing_search.py

@@ -6,6 +6,7 @@ from bs4 import BeautifulSoup
 from app.logger import logger
 from app.tool.search.base import WebSearchEngine
 
+
 ABSTRACT_MAX_LENGTH = 300
 
 USER_AGENTS = [

app/tool/tool_collection.py

@@ -8,6 +8,9 @@ from app.tool.base import BaseTool, ToolFailure, ToolResult
 class ToolCollection:
     """A collection of defined tools."""
 
+    class Config:
+        arbitrary_types_allowed = True
+
     def __init__(self, *tools: BaseTool):
         self.tools = tools
         self.tool_map = {tool.name: tool for tool in tools}

app/tool/web_search.py

@@ -4,13 +4,14 @@ from typing import List
 from tenacity import retry, stop_after_attempt, wait_exponential
 
 from app.config import config
+from app.logger import logger
 from app.tool.base import BaseTool
 from app.tool.search import (
     BaiduSearchEngine,
     BingSearchEngine,
     DuckDuckGoSearchEngine,
     GoogleSearchEngine,
-    WebSearchEngine
+    WebSearchEngine,
 )
 
 
@@ -44,6 +45,8 @@ class WebSearch(BaseTool):
     async def execute(self, query: str, num_results: int = 10) -> List[str]:
         """
         Execute a Web search and return a list of URLs.
+        Tries engines in order based on configuration, falling back if an engine fails with errors.
+        If all engines fail, it will wait and retry up to the configured number of times.
 
         Args:
             query (str): The search query to submit to the search engine.
@@ -52,37 +55,109 @@ class WebSearch(BaseTool):
         Returns:
             List[str]: A list of URLs matching the search query.
         """
+        # Get retry settings from config
+        retry_delay = 60  # Default to 60 seconds
+        max_retries = 3  # Default to 3 retries
+
+        if config.search_config:
+            retry_delay = getattr(config.search_config, "retry_delay", 60)
+            max_retries = getattr(config.search_config, "max_retries", 3)
+
+        # Try searching with retries when all engines fail
+        for retry_count in range(
+            max_retries + 1
+        ):  # +1 because first try is not a retry
+            links = await self._try_all_engines(query, num_results)
+            if links:
+                return links
+
+            if retry_count < max_retries:
+                # All engines failed, wait and retry
+                logger.warning(
+                    f"All search engines failed. Waiting {retry_delay} seconds before retry {retry_count + 1}/{max_retries}..."
+                )
+                await asyncio.sleep(retry_delay)
+            else:
+                logger.error(
+                    f"All search engines failed after {max_retries} retries. Giving up."
+                )
+
+        return []
+
+    async def _try_all_engines(self, query: str, num_results: int) -> List[str]:
+        """
+        Try all search engines in the configured order.
+
+        Args:
+            query (str): The search query to submit to the search engine.
+            num_results (int): The number of search results to return.
+
+        Returns:
+            List[str]: A list of URLs matching the search query, or empty list if all engines fail.
+        """
         engine_order = self._get_engine_order()
+        failed_engines = []
+
         for engine_name in engine_order:
             engine = self._search_engine[engine_name]
             try:
+                logger.info(f"🔎 Attempting search with {engine_name.capitalize()}...")
                 links = await self._perform_search_with_engine(
                     engine, query, num_results
                 )
                 if links:
+                    if failed_engines:
+                        logger.info(
+                            f"Search successful with {engine_name.capitalize()} after trying: {', '.join(failed_engines)}"
+                        )
                     return links
             except Exception as e:
-                print(f"Search engine '{engine_name}' failed with error: {e}")
+                failed_engines.append(engine_name.capitalize())
+                is_rate_limit = "429" in str(e) or "Too Many Requests" in str(e)
+
+                if is_rate_limit:
+                    logger.warning(
+                        f"⚠️ {engine_name.capitalize()} search engine rate limit exceeded, trying next engine..."
+                    )
+                else:
+                    logger.warning(
+                        f"⚠️ {engine_name.capitalize()} search failed with error: {e}"
+                    )
+
+        if failed_engines:
+            logger.error(f"All search engines failed: {', '.join(failed_engines)}")
         return []
 
     def _get_engine_order(self) -> List[str]:
         """
         Determines the order in which to try search engines.
-        Preferred engine is first (based on configuration), followed by the remaining engines.
+        Preferred engine is first (based on configuration), followed by fallback engines,
+        and then the remaining engines.
 
         Returns:
             List[str]: Ordered list of search engine names.
         """
         preferred = "google"
-        if config.search_config and config.search_config.engine:
+        fallbacks = []
-            preferred = config.search_config.engine.lower()
+
+        if config.search_config:
+            if config.search_config.engine:
+                preferred = config.search_config.engine.lower()
+            if config.search_config.fallback_engines:
+                fallbacks = [
+                    engine.lower() for engine in config.search_config.fallback_engines
+                ]
 
         engine_order = []
+        # Add preferred engine first
         if preferred in self._search_engine:
             engine_order.append(preferred)
-        for key in self._search_engine:
+
-            if key not in engine_order:
+        # Add configured fallback engines in order
-                engine_order.append(key)
+        for fallback in fallbacks:
+            if fallback in self._search_engine and fallback not in engine_order:
+                engine_order.append(fallback)
+
         return engine_order
 
     @retry(

config/config.example.toml

@@ -6,6 +6,14 @@ api_key = "YOUR_API_KEY"                    # Your API key
 max_tokens = 8192                           # Maximum number of tokens in the response
 temperature = 0.0                           # Controls randomness
 
+# [llm] # Amazon Bedrock
+# api_type = "aws"                                       # Required
+# model = "us.anthropic.claude-3-7-sonnet-20250219-v1:0" # Bedrock supported modelID
+# base_url = "bedrock-runtime.us-west-2.amazonaws.com"   # Not used now
+# max_tokens = 8192
+# temperature = 1.0
+# api_key = "bear"                                       # Required but not used for Bedrock
+
 # [llm] #AZURE OPENAI:
 # api_type= 'azure'
 # model = "YOUR_MODEL_NAME" #"gpt-4o-mini"
@@ -65,6 +73,13 @@ temperature = 0.0                           # Controls randomness for vision mod
 # [search]
 # Search engine for agent to use. Default is "Google", can be set to "Baidu" or "DuckDuckGo".
 #engine = "Google"
+# Fallback engine order. Default is ["DuckDuckGo", "Baidu"] - will try in this order after primary engine fails.
+#fallback_engines = ["DuckDuckGo", "Baidu"]
+# Seconds to wait before retrying all engines again when they all fail due to rate limits. Default is 60.
+#retry_delay = 60
+# Maximum number of times to retry all engines when all fail. Default is 3.
+#max_retries = 3
+
 
 ## Sandbox configuration
 #[sandbox]

requirements.txt

@@ -27,3 +27,9 @@ playwright~=1.50.0
 docker~=7.1.0
 pytest~=8.3.5
 pytest-asyncio~=0.25.3
+
+mcp~=1.4.1
+httpx>=0.27.0
+tomli>=2.0.0
+
+boto3~=1.37.16

run_mcp.py

@@ -0,0 +1,116 @@
+#!/usr/bin/env python
+import argparse
+import asyncio
+import sys
+
+from app.agent.mcp import MCPAgent
+from app.config import config
+from app.logger import logger
+
+
+class MCPRunner:
+    """Runner class for MCP Agent with proper path handling and configuration."""
+
+    def __init__(self):
+        self.root_path = config.root_path
+        self.server_reference = "app.mcp.server"
+        self.agent = MCPAgent()
+
+    async def initialize(
+        self,
+        connection_type: str,
+        server_url: str | None = None,
+    ) -> None:
+        """Initialize the MCP agent with the appropriate connection."""
+        logger.info(f"Initializing MCPAgent with {connection_type} connection...")
+
+        if connection_type == "stdio":
+            await self.agent.initialize(
+                connection_type="stdio",
+                command=sys.executable,
+                args=["-m", self.server_reference],
+            )
+        else:  # sse
+            await self.agent.initialize(connection_type="sse", server_url=server_url)
+
+        logger.info(f"Connected to MCP server via {connection_type}")
+
+    async def run_interactive(self) -> None:
+        """Run the agent in interactive mode."""
+        print("\nMCP Agent Interactive Mode (type 'exit' to quit)\n")
+        while True:
+            user_input = input("\nEnter your request: ")
+            if user_input.lower() in ["exit", "quit", "q"]:
+                break
+            response = await self.agent.run(user_input)
+            print(f"\nAgent: {response}")
+
+    async def run_single_prompt(self, prompt: str) -> None:
+        """Run the agent with a single prompt."""
+        await self.agent.run(prompt)
+
+    async def run_default(self) -> None:
+        """Run the agent in default mode."""
+        prompt = input("Enter your prompt: ")
+        if not prompt.strip():
+            logger.warning("Empty prompt provided.")
+            return
+
+        logger.warning("Processing your request...")
+        await self.agent.run(prompt)
+        logger.info("Request processing completed.")
+
+    async def cleanup(self) -> None:
+        """Clean up agent resources."""
+        await self.agent.cleanup()
+        logger.info("Session ended")
+
+
+def parse_args() -> argparse.Namespace:
+    """Parse command line arguments."""
+    parser = argparse.ArgumentParser(description="Run the MCP Agent")
+    parser.add_argument(
+        "--connection",
+        "-c",
+        choices=["stdio", "sse"],
+        default="stdio",
+        help="Connection type: stdio or sse",
+    )
+    parser.add_argument(
+        "--server-url",
+        default="http://127.0.0.1:8000/sse",
+        help="URL for SSE connection",
+    )
+    parser.add_argument(
+        "--interactive", "-i", action="store_true", help="Run in interactive mode"
+    )
+    parser.add_argument("--prompt", "-p", help="Single prompt to execute and exit")
+    return parser.parse_args()
+
+
+async def run_mcp() -> None:
+    """Main entry point for the MCP runner."""
+    args = parse_args()
+    runner = MCPRunner()
+
+    try:
+        await runner.initialize(args.connection, args.server_url)
+
+        if args.prompt:
+            await runner.run_single_prompt(args.prompt)
+        elif args.interactive:
+            await runner.run_interactive()
+        else:
+            await runner.run_default()
+
+    except KeyboardInterrupt:
+        logger.info("Program interrupted by user")
+    except Exception as e:
+        logger.error(f"Error running MCPAgent: {str(e)}", exc_info=True)
+        sys.exit(1)
+    finally:
+        await runner.cleanup()
+
+
+if __name__ == "__main__":
+    asyncio.run(run_mcp())

run_mcp_server.py

@@ -0,0 +1,11 @@
+# coding: utf-8
+# A shortcut to launch OpenManus MCP server, where its introduction also solves other import issues.
+from app.mcp.server import MCPServer, parse_args
+
+
+if __name__ == "__main__":
+    args = parse_args()
+
+    # Create and run server (maintaining original flow)
+    server = MCPServer()
+    server.run(transport=args.transport)