universal-mcp-agents 0.1.17__py3-none-any.whl → 0.1.19rc1__py3-none-any.whl

universal_mcp/agents/__init__.py

@@ -7,6 +7,7 @@ from universal_mcp.agents.codeact import CodeActAgent as CodeActScript
 from universal_mcp.agents.codeact0 import CodeActPlaybookAgent as CodeActRepl
 from universal_mcp.agents.react import ReactAgent
 from universal_mcp.agents.simple import SimpleAgent
+from universal_mcp.agents.unified import UnifiedAgent
 
 
 def get_agent(agent_name: Literal["react", "simple", "builder", "bigtool", "codeact-script", "codeact-repl"]):
@@ -22,6 +23,8 @@ def get_agent(agent_name: Literal["react", "simple", "builder", "bigtool", "code
         return CodeActScript
     elif agent_name == "codeact-repl":
         return CodeActRepl
+    elif agent_name == "unified":
+        return UnifiedAgent
     else:
         raise ValueError(
             f"Unknown agent: {agent_name}. Possible values:  react, simple, builder, bigtool, codeact-script, codeact-repl"

universal_mcp/agents/codeact0/__main__.py

@@ -19,12 +19,6 @@ async def main():
         memory=memory,
     )
     print("Starting agent...")
-    # await agent.ainit()
-    # await agent.run_interactive()
-    # async for event in agent.stream(
-    #     user_input="Fetch unsubscribe links from my Gmail inbox for promo emails I have received in the last 7 days"
-    # ):
-    #     print(event.content, end="")
     result = await agent.invoke(
         user_input="Fetch unsubscribe links from my Gmail inbox for promo emails I have received in the last 7 days"
     )

universal_mcp/agents/codeact0/llm_tool.py

@@ -27,7 +27,7 @@ def smart_print(data: Any) -> None:
     Args:
         data: Either a dictionary with string keys, or a list of such dictionaries
     """
-    print(light_copy(data))  # noqa
+    print(light_copy(data))  # noqa: T201
 
 
 def creative_writer(
@@ -275,105 +275,3 @@ def data_extractor(
         .invoke(prompt)
     )
     return cast(dict[str, Any], response)
-
-
-# news_articles_schema = {
-#     "type": "object",
-#     "properties": {
-#         "articles": {
-#             "type": "array",
-#             "title": "Articles",
-#             "description": "List of news articles",
-#             "items": {
-#                 "type": "object",
-#                 "properties": {
-#                     "headline": {
-#                         "type": "string",
-#                         "title": "Headline",
-#                         "description": "The headline of the news article"
-#                     },
-#                     "url": {
-#                         "type": "string",
-#                         "title": "URL",
-#                         "description": "The URL of the news article"
-#                     }
-#                 },
-#                 "required": ["headline", "url"],
-#             }
-#         }
-#     },
-#     "required": ["articles"],
-# }
-
-
-# news_articles_schema = {
-#   "title": "NewsArticleList",
-#   "description": "A list of news articles with headlines and URLs",
-#   "type": "object",
-#   "properties": {
-#     "articles": {
-#       "type": "array",
-#       "items": {
-#         "type": "object",
-#         "properties": {
-#           "headline": {
-#             "type": "string"
-#           },
-#           "url": {
-#             "type": "string"
-#           }
-#         },
-#         "required": ["headline", "url"]
-#       }
-#     }
-#   },
-#   "required": ["articles"]
-# }
-# model = init_chat_model(model="claude-4-sonnet-20250514", temperature=0)
-# structured_model = model.with_structured_output(news_articles_schema)
-
-
-# class TwitterComment(BaseModel):
-#     skip: bool
-#     reason: str
-#     comment: str
-
-# twitter_comment_schema = {
-#   "title": "TwitterComment",
-#   "description": "A twitter comment to engage with followers",
-#   "type": "object",
-#   "properties": {
-#     "skip": {
-#       "type": "boolean"
-#     },
-#     "reason": {
-#       "type": "string"
-#     },
-#     "comment": {
-#       "type": "string"
-#     },
-#     "tagged_profiles": {
-#       "type": "array",
-#       "items": {
-#         "type": "string"
-#       }
-#     }
-#   },
-#   "required": ["skip", "reason"]
-# }
-
-# comment = {
-#     "tweet_id": "08109402",
-#     "handle": "@iamnishant",
-#     "text": "Hey really loved this tweet! Well said 💯"
-# }
-
-# comment_instructions = (
-#     "Goal is to engage with my twitter followers who have commented on my tweets."
-#     "Please generate a single line, context-aware, conversational reply for the given comment."
-#     "- Use social media language (can use hinglish)."
-#     "- Skip the reply, if the comment is too generic."
-#     "- Also tag relevant people in the reply."
-# )
-
-# my_reply = call_llm(comment_instructions, comment, twitter_comment_schema)

universal_mcp/agents/codeact0/playbook_agent.py

@@ -2,18 +2,16 @@ import inspect
 import json
 import re
 from collections.abc import Callable
-from dataclasses import dataclass
-from pathlib import Path
 from typing import Literal, cast
 
-from langchain_core.messages import AIMessage, RemoveMessage, ToolMessage
+from langchain_core.messages import AIMessage, ToolMessage
 from langchain_core.tools import StructuredTool
 from langchain_core.tools import tool as create_tool
 from langgraph.checkpoint.base import BaseCheckpointSaver
 from langgraph.graph import START, StateGraph
 from langgraph.types import Command, RetryPolicy
 from universal_mcp.tools.registry import ToolRegistry
-from universal_mcp.types import ToolFormat, ToolConfig
+from universal_mcp.types import ToolConfig, ToolFormat
 
 from universal_mcp.agents.base import BaseAgent
 from universal_mcp.agents.codeact0.llm_tool import ai_classify, call_llm, data_extractor, smart_print
@@ -22,10 +20,14 @@ from universal_mcp.agents.codeact0.prompts import (
 )
 from universal_mcp.agents.codeact0.sandbox import eval_unsafe, execute_ipython_cell
 from universal_mcp.agents.codeact0.state import CodeActState
-from universal_mcp.agents.codeact0.tools import create_meta_tools, enter_playbook_mode, exit_playbook_mode, get_valid_tools
+from universal_mcp.agents.codeact0.tools import (
+    create_meta_tools,
+    enter_playbook_mode,
+    get_valid_tools,
+)
 from universal_mcp.agents.codeact0.utils import inject_context, smart_truncate
 from universal_mcp.agents.llm import load_chat_model
-from universal_mcp.agents.utils import filter_retry_on, get_message_text, convert_tool_ids_to_dict
+from universal_mcp.agents.utils import convert_tool_ids_to_dict, filter_retry_on, get_message_text
 
 PLAYBOOK_PLANNING_PROMPT = """Now, you are tasked with creating a reusable playbook from the user's previous workflow.
 
@@ -48,7 +50,6 @@ Example:
 Now create a plan based on the conversation history. Enclose it between ``` and ```. Ask the user if the plan is okay."""
 
 
-
 PLAYBOOK_CONFIRMING_PROMPT = """Now, you are tasked with confirming the playbook plan. Return True if the user is happy with the plan, False otherwise. Do not say anything else in your response. The user response will be the last message in the chain.
 """
 
@@ -80,7 +81,7 @@ class CodeActPlaybookAgent(BaseAgent):
             memory=memory,
             **kwargs,
         )
-        self.model_instance = load_chat_model(model, thinking=True)
+        self.model_instance = load_chat_model(model)
         self.tools_config = tools or []
         self.registry = registry
         self.playbook_registry = playbook_registry
@@ -92,20 +93,19 @@ class CodeActPlaybookAgent(BaseAgent):
         meta_tools = create_meta_tools(self.registry)
         additional_tools = [smart_print, data_extractor, ai_classify, call_llm, meta_tools["web_search"]]
         self.additional_tools = [t if isinstance(t, StructuredTool) else create_tool(t) for t in additional_tools]
+
         async def call_model(state: CodeActState) -> Command[Literal["sandbox", "execute_tools"]]:
             self.exported_tools = []
             if self.tools_config:
                 # Convert dict format to list format if needed
                 if isinstance(self.tools_config, dict):
                     self.tools_config = [
-                        f"{provider}__{tool}" 
-                        for provider, tools in self.tools_config.items() 
-                        for tool in tools
+                        f"{provider}__{tool}" for provider, tools in self.tools_config.items() for tool in tools
                     ]
                 if not self.registry:
                     raise ValueError("Tools are configured but no registry is provided")
                 # Langchain tools are fine
-            self.tools_config.extend(state.get('selected_tool_ids',[]))
+            self.tools_config.extend(state.get("selected_tool_ids", []))
             self.exported_tools = await self.registry.export_tools(self.tools_config, ToolFormat.LANGCHAIN)
             self.final_instructions, self.tools_context = create_default_prompt(
                 self.exported_tools, self.additional_tools, self.instructions
@@ -167,7 +167,7 @@ class CodeActPlaybookAgent(BaseAgent):
                         )
                         return Command(
                             goto="playbook",
-                            update={"playbook_mode": "planning", "messages": [tool_message]}, #Entered Playbook mode
+                            update={"playbook_mode": "planning", "messages": [tool_message]},  # Entered Playbook mode
                         )
                     elif tool_call["name"] == "execute_ipython_cell":
                         return Command(goto="sandbox")
@@ -261,13 +261,12 @@ class CodeActPlaybookAgent(BaseAgent):
                 response = cast(AIMessage, response)
                 response_text = get_message_text(response)
                 # Extract plan from response text between triple backticks
-                plan_match = re.search(r'```(.*?)```', response_text, re.DOTALL)
+                plan_match = re.search(r"```(.*?)```", response_text, re.DOTALL)
                 if plan_match:
                     plan = plan_match.group(1).strip()
                 else:
                     plan = response_text.strip()
                 return Command(update={"messages": [response], "playbook_mode": "confirming", "plan": plan})
-                
 
             elif playbook_mode == "confirming":
                 confirmation_instructions = self.instructions + PLAYBOOK_CONFIRMING_PROMPT
@@ -279,8 +278,6 @@ class CodeActPlaybookAgent(BaseAgent):
                 else:
                     return Command(goto="playbook", update={"playbook_mode": "planning"})
 
-                    
-
             elif playbook_mode == "generating":
                 generating_instructions = self.instructions + PLAYBOOK_GENERATING_PROMPT
                 messages = [{"role": "system", "content": generating_instructions}] + state["messages"]
@@ -327,25 +324,19 @@ class CodeActPlaybookAgent(BaseAgent):
                     saved_note = f"Failed to save generated playbook as Agent '{function_name}': {e}"
 
                 # Mock tool call for exit_playbook_mode (for testing/demonstration)
-                mock_exit_tool_call = {
-                    "name": "exit_playbook_mode",
-                    "args": {},
-                    "id": "mock_exit_playbook_123"
-                }
-                mock_assistant_message = AIMessage(
-                    content=saved_note, 
-                    tool_calls=[mock_exit_tool_call]
-                )
+                mock_exit_tool_call = {"name": "exit_playbook_mode", "args": {}, "id": "mock_exit_playbook_123"}
+                mock_assistant_message = AIMessage(content=saved_note, tool_calls=[mock_exit_tool_call])
 
-                
                 # Mock tool response for exit_playbook_mode
                 mock_exit_tool_response = ToolMessage(
                     content=json.dumps(f"Exited Playbook Mode.{saved_note}"),
                     name="exit_playbook_mode",
-                    tool_call_id="mock_exit_playbook_123"
+                    tool_call_id="mock_exit_playbook_123",
                 )
 
-                return Command(update={"messages": [mock_assistant_message, mock_exit_tool_response], "playbook_mode": "normal"})
+                return Command(
+                    update={"messages": [mock_assistant_message, mock_exit_tool_response], "playbook_mode": "normal"}
+                )
 
         def route_entry(state: CodeActState) -> Literal["call_model", "playbook"]:
             """Route to either normal mode or playbook creation"""

universal_mcp/agents/codeact0/prompts.py

@@ -9,102 +9,38 @@ from universal_mcp.agents.codeact0.utils import schema_to_signature
 uneditable_prompt = """
 You are **Wingmen**, an AI Assistant created by AgentR — a creative, straight-forward, and direct principal software engineer with access to tools.
 
-## Responsibilities
-
-- **Answer directly** if the task is simple (e.g. print, math, general knowledge).
-- For any task requiring logic, execution, or data handling, use `execute_ipython_cell`.
-- For writing or NLP tasks (summarizing, generating, extracting), always use AI functions via code — never respond directly.
-
-## Tool vs. Function: Required Separation
-
-You must clearly distinguish between tools (called via the tool calling API) and internal functions (used inside code blocks).
-
-### Tools — Must Be Called via Tool Calling API
-
-These must be called using **tool calling**, not from inside code blocks:
-
-- `execute_ipython_cell` — For running any Python code or logic.
-- `search_functions` — To discover available functions for a task.
-- `load_functions` — To load a specific function by full ID.
-
-**Do not attempt to call these inside `python` code.**
-Use tool calling syntax for these operations.
-
-### Functions — Must Be Used Inside Code Blocks
-
-All other functions, including LLM functions, must always be used within code executed by `execute_ipython_cell`. These include:
-
-- `smart_print()` — For inspecting unknown data structures before looping.
-- `asyncio.run()` — For wrapping and executing asynchronous logic. You must not use await outside an async function. And the async function must be called by `asyncio.run()`.
-- Any functions for applications loaded via `load_functions`.
-- Any logic, data handling, writing, NLP, generation, summarization, or extraction functionality of LLMs.
-
-These must be called **inside a Python code block**, and that block must be executed using `execute_ipython_cell`.
-
-## Tool/Function Usage Policy
-
-1. **Always Use Tools/Functions for Required Tasks**
-   Any searching, loading, or executing must be done using a tool/function call. Never answer manually if a tool/function is appropriate.
-
-2. **Use Existing Functions First**
-   Use existing functions if available. Otherwise, use `search_functions` with a concise query describing the task.
-
-3. **Load Only Relevant Tools**
-   When calling `load_functions`, include only relevant function IDs.
-   - Prefer connected applications over unconnected ones.
-   - If multiple functions match (i.e. if none are connected, or multiple are connected), ask the user to choose.
-   - After loading a tool, you do not need to import/declare it again. It can be called directly in further cells.
-
-4. **Follow First Turn Process Strictly**
-   On the **first turn**, do only **one** of the following:
-   - Handle directly (if trivial)
-   - Use a tool/function (`execute_ipython_cell`, `search_functions`, etc.)
-
-   **Do not extend the conversation on the first message.**
-
-## Coding Rules
-
-- Use `smart_print()` to inspect unknown structures, especially those received from function outputs, before looping or branching.
-- Validate logic with a single item before processing lists or large inputs.
-- Try to achieve as much as possible in a single code block.
-- Use only pre-installed Python libraries. Do import them once before using.
-- Outer level functions, variables, classes, and imports declared previously can be used in later cells.
-- For all functions, call using keyword arguments only. DO NOT use any positional arguments.
-
-### **Async Function Usage — Critical**
-
-When calling asynchronous functions:
-- You must define or use an **inner async function**.
-- Use `await` only **inside** that async function.
-- Run it using `asyncio.run(<function_name>())` **without** `await` at the outer level.
-
-**Wrong - Using `await` outside an async function**
-```
-result = await some_async_function()
-```
-**Wrong  - Attaching await before asyncio.run**.
-`await asyncio.run(main())`
-These will raise SyntaxError: 'await' outside async function
-The correct method is the following-
-```
-import asyncio
-async def some_async_function():
-    ...
-
-async def main():
-    result = await some_async_function()
-    print(result)
-
-asyncio.run(main())
-#or
-result = asyncio.run(some_async_function(arg1 = <arg1>))
-```
-## Output Formatting
-- All code results must be returned in **Markdown**.
-- The user cannot see raw output, so format results clearly:
-    - Use tables for structured data.
-    - Provide links for files or images.
-    - Be explicit in formatting to ensure readability.
+Your job is to answer the user's question or perform the task they ask for.
+- Answer simple questions (which do not require you to write any code or access any external resources) directly. Note that any operation that involves using ONLY print functions should be answered directly.
+- For task requiring operations or access to external resources, you should achieve the task by executing Python code snippets.
+- You have access to `execute_ipython_cell` tool that allows you to execute Python code in an IPython notebook cell.
+- You also have access to two tools for finding and loading more python functions- `search_functions` and `load_functions`, which you must use for finding functions for using different external applications. Prefer pre-loaded or functions already available when possible, and prioritize connected applications over unconnected ones. When this is not enough to break a tie between similar applications, ask the user.
+- In writing or natural language processing tasks DO NOT answer directly. Instead use `execute_ipython_cell` tool with the AI functions provided to you for tasks like summarizing, text generation, classification, data extraction from text or unstructured data, etc. Avoid hardcoded approaches to classification, data extraction.
+- The code you write will be executed in a sandbox environment, and you can use the output of previous executions in your code. variables, functions, imports are retained.
+- Read and understand the output of the previous code snippet and use it to answer the user's request. Note that the code output is NOT visible to the user, so after the task is complete, you have to give the output to the user in a markdown format.
+- If needed, feel free to ask for more information from the user (without using the `execute_ipython_cell` tool) to clarify the task.
+
+GUIDELINES for writing code:
+- Variables defined at the top level of previous code snippets can be referenced in your code.
+- External functions which return a dict or list[dict] are ambiguous. Therefore, you MUST explore the structure of the returned data using `smart_print()` statements before using it, printing keys and values. `smart_print` truncates long strings from data, preventing huge output logs.
+- When an operation involves running a fixed set of steps on a list of items, run one run correctly and then use a for loop to run the steps on each item in the list.
+- In a single code snippet, try to achieve as much as possible.
+- You can only import libraries that come pre-installed with Python. For external functions, use the search and load tools to access them in the code.
+- For displaying final results to the user, you must present your output in markdown format, including image links, so that they are rendered and displayed to the user. The code output is NOT visible to the user.
+- Call all functions using keyword arguments only, never positional arguments.
+- Async Functions (Critical): Use them only as follows-
+Case 1: Top-level await without asyncio.run()
+    Wrap in async function and call with asyncio.run():
+    async def main():
+        result = await some_async_function()
+        return result
+    asyncio.run(main())
+Case 2: Using asyncio.run() directly
+If code already contains asyncio.run(), use as-is — do not wrap again:
+    asyncio.run(some_async_function())
+Rules:
+- Never use await outside an async function
+- Never use await asyncio.run()
+- Never nest asyncio.run() calls
 """
 
 

universal_mcp/agents/codeact0/tools.py

@@ -6,18 +6,19 @@ from langchain_core.tools import tool
 from universal_mcp.tools.registry import ToolRegistry
 from universal_mcp.types import ToolFormat
 
-MAX_LENGHT=100
+MAX_LENGHT = 100
+
 
 def enter_playbook_mode():
     """Call this function to enter playbook mode. Playbook mode is when the user wants to store a repeated task as a script with some inputs for the future."""
     return
 
+
 def exit_playbook_mode():
     """Call this function to exit playbook mode. Playbook mode is when the user wants to store a repeated task as a script with some inputs for the future."""
     return
 
 
-
 def create_meta_tools(tool_registry: ToolRegistry) -> dict[str, Any]:
     """Create the meta tools for searching and loading tools"""
 
@@ -46,7 +47,7 @@ def create_meta_tools(tool_registry: ToolRegistry) -> dict[str, Any]:
                 for tool in tools_list:
                     app = tool["id"].split("__")[0]
                     tool_id = tool["id"]
-                    
+
                     # Check if within limit and add to set (automatically deduplicates)
                     if len(app_tools[app]) < MAX_LENGTH:
                         cleaned_desc = tool["description"].split("Context:")[0].strip()
@@ -103,29 +104,30 @@ def create_meta_tools(tool_registry: ToolRegistry) -> dict[str, Any]:
         return f"Successfully loaded {len(tool_ids)} functions: {tool_ids}"
 
     @tool
-    async def web_search(query: str) -> list:
-        """Search the web for the given query and return structured search results.
+    async def web_search(query: str) -> dict:
+        """
+        Get an LLM answer to a question informed by Exa search results.
 
-        Do not use for app-specific searches (for example, reddit or linkedin searches
+        This tool performs an Exa `/answer` request, which:
+        1. Provides a **direct answer** for factual queries (e.g., "What is the capital of France?" → "Paris")
+        2. Generates a **summary with citations** for open-ended questions
+        (e.g., "What is the state of AI in healthcare?" → A detailed summary with source links)
 
+        Args:
+            query (str): The question or topic to answer.
         Returns:
-            list: A list of up to 10 search result dictionaries, each containing:
-                - id (str): Unique identifier, typically the URL
-                - title (str): The title/headline of the search result
-                - url (str): The web URL of the result
-                - publishedDate (str): ISO 8601 formatted date (e.g., "2025-01-01T00:00:00.000Z")
-                - author (str): Author name (may be empty string)
-                - summary (str): Text summary/snippet of the content
-                - image (str): URL to associated image (if available)
-
-        Example:
-            results = await web_search(query="python programming")
+            dict: A structured response containing only:
+                - answer (str): Generated answer
+                - citations (list[dict]): List of cited sources
         """
-        await tool_registry.export_tools(["exa__search_with_filters"], ToolFormat.LANGCHAIN)
-        response = await tool_registry.call_tool(
-            "exa__search_with_filters", {"query": query, "contents": {"summary": True}}
-        )
-        return response["results"]
+        await tool_registry.export_tools(["exa__answer"], ToolFormat.LANGCHAIN)
+        response = await tool_registry.call_tool("exa__answer", {"query": query, "text": True})
+
+        # Extract only desired fields
+        return {
+            "answer": response.get("answer"),
+            "citations": response.get("citations", []),
+        }
 
     return {"search_functions": search_functions, "load_functions": load_functions, "web_search": web_search}
 

universal_mcp/agents/codeact0/utils.py

@@ -6,7 +6,7 @@ from typing import Any
 
 from langchain_core.messages import BaseMessage
 
-MAX_CHARS = 700
+MAX_CHARS = 5000
 
 
 def light_copy(data):

universal_mcp/agents/sandbox.py

@@ -0,0 +1,90 @@
+import contextlib
+import inspect
+import io
+import queue
+import re
+import socket
+import threading
+import types
+from typing import Any
+
+from universal_mcp.agents.codeact0.utils import derive_context
+
+
+class Sandbox:
+    """
+    A class to execute code safely in a sandboxed environment with a timeout.
+    """
+
+    def __init__(self, timeout: int = 180):
+        """
+        Initializes the Sandbox.
+        Args:
+            timeout: The timeout for code execution in seconds.
+        """
+        self.timeout = timeout
+        self._locals: dict[str, Any] = {}
+        self.add_context: dict[str, Any] = {}
+
+    def run(self, code: str) -> tuple[str, dict[str, Any], dict[str, Any]]:
+        """
+        Execute code safely with a timeout.
+        - Returns (output_str, filtered_locals_dict, new_add_context)
+        - Errors or timeout are returned as output_str.
+        - Previous variables in _locals persist across calls.
+        """
+
+        EXCLUDE_TYPES = (
+            types.ModuleType,
+            type(re.match("", "")),
+            type(threading.Lock()),
+            type(threading.RLock()),
+            threading.Event,
+            threading.Condition,
+            threading.Semaphore,
+            queue.Queue,
+            socket.socket,
+            io.IOBase,
+        )
+
+        result_container = {"output": "<no output>"}
+
+        def target():
+            try:
+                with contextlib.redirect_stdout(io.StringIO()) as f:
+                    exec(code, self._locals, self._locals)
+                result_container["output"] = f.getvalue() or "<code ran, no output printed to stdout>"
+            except Exception as e:
+                result_container["output"] = "Error during execution: " + str(e)
+
+        thread = threading.Thread(target=target)
+        thread.start()
+        thread.join(self.timeout)
+
+        if thread.is_alive():
+            result_container["output"] = f"Code timeout: code execution exceeded {self.timeout} seconds."
+
+        # Filter locals for picklable/storable variables
+        all_vars = {}
+        for key, value in self._locals.items():
+            if key == "__builtins__":
+                continue
+            if inspect.iscoroutine(value) or inspect.iscoroutinefunction(value):
+                continue
+            if inspect.isasyncgen(value) or inspect.isasyncgenfunction(value):
+                continue
+            if isinstance(value, EXCLUDE_TYPES):
+                continue
+            if not callable(value) or not hasattr(value, "__name__"):
+                all_vars[key] = value
+
+        self._locals = all_vars
+
+        # Safely derive context
+        try:
+            self.add_context = derive_context(code, self.add_context)
+        except Exception:
+            # Keep the old context if derivation fails
+            pass
+
+        return result_container["output"], self._locals, self.add_context

universal_mcp/agents/unified/README.md

@@ -0,0 +1,45 @@
+# Unified Agent
+
+The Unified Agent is a sophisticated AI assistant designed to understand and execute tasks by writing and running Python code. It operates within a secure sandbox environment and can leverage a variety of tools to interact with external systems and perform complex operations. A key feature of the Unified Agent is its ability to create reusable "playbooks" from user workflows, enabling automation of repeated tasks.
+
+## Architecture
+
+The agent's architecture is built upon the LangGraph library, creating a state machine that cycles between thinking (calling a language model) and acting (executing code or tools).
+
+### Core Components:
+
+*   **`UnifiedAgent`**: The fundamental agent implementation. It processes user requests, writes Python code, and executes it in a sandbox to achieve the desired outcome. It also has a "playbook mode" to generate reusable Python functions from a user's workflow.
+*   **State Graph (`CodeActState`)**: The agent's logic is defined as a state graph. The primary nodes are:
+    *   `call_model`: Invokes the language model to generate Python code or select a tool based on the current state and user input.
+    *   `sandbox`: Executes the generated Python code using a safe `eval` function with a timeout. The results and any errors are captured and fed back into the state.
+    *   `execute_tools`: Handles the execution of meta-tools for searching, loading, and interacting with external functions.
+    *   `playbook`: Manages the playbook creation process, including planning, user confirmation, and code generation.
+*   **Sandbox (`sandbox.py`)**: A secure execution environment that runs Python code in a separate thread with a timeout. It ensures that the agent's code execution is isolated and cannot harm the host system.
+*   **Tools**: The agent has access to a set of powerful tools:
+    *   `execute_ipython_cell`: The primary tool for executing arbitrary Python code snippets.
+    *   **AI Functions (`llm_tool.py`)**: A suite of functions (`generate_text`, `classify_data`, `extract_data`, `call_llm`) that allow the agent to delegate complex reasoning, classification, and data extraction tasks to a language model.
+    *   **Meta Tools (`tools.py`)**: Functions like `search_functions` and `load_functions` that enable the agent to dynamically discover and load new tools from a `ToolRegistry`.
+
+## Playbook Mode
+
+A key feature of the Unified Agent is its ability to create reusable "playbooks". When a user performs a task that they might want to repeat in the future, they can trigger the playbook mode. The agent will then:
+
+1.  **Plan:** Analyze the workflow and create a step-by-step plan for a reusable function, identifying user-specific variables that should become function parameters.
+2.  **Confirm:** Ask the user for confirmation of the generated plan.
+3.  **Generate:** Generate a Python function based on the confirmed plan. This function can be saved and executed later to automate the task.
+
+## Getting Started (`__main__.py`)
+
+The `__main__.py` file serves as a simple command-line interface for interacting with the agent. It demonstrates how to instantiate the `UnifiedAgent`, configure it with tools, and invoke it with a user request. This allows for easy testing and experimentation with the agent's capabilities.
+
+To run the agent, execute the following command from the root of the repository:
+```bash
+uv run python -m src.universal_mcp.agents.unified.__main__
+```
+
+Major TODO:
+- [] Improve LLM Tools
+    - [] Use smaller dedicated models for universal_write, clasify etc
+- Improve Sandbox
+    - [] Support saving loading context
+    - [] Direct async tool support

universal_mcp/agents/unified/__init__.py

@@ -0,0 +1,3 @@
+from .agent import UnifiedAgent
+
+__all__ = ["UnifiedAgent"]