mcp-use__mcp-use/docs/python/agent/agent-configuration.mdx

---
title: "Agent Configuration"
description: "Configure MCPAgent behavior and LLM integration"
icon: "cog"
---

<Info>
This guide covers MCPAgent configuration options for customizing agent behavior and LLM integration. For client configuration, see the [Client Configuration](/python/client/client-configuration) guide.
</Info>

## API Keys

<Warning>
Never hardcode API keys in your source code. Always use environment variables for security.
</Warning>

Since agents use LLM providers that require API keys, you need to configure them properly:

<Tabs>
  <Tab title=".env File (Recommended)">
    Create a `.env` file in your project root:

    ```bash .env
    # OpenAI
    OPENAI_API_KEY=your_api_key_here
    # Anthropic
    ANTHROPIC_API_KEY=your_api_key_here
    # Groq
    GROQ_API_KEY=your_api_key_here
    # Google
    GOOGLE_API_KEY=your_api_key_here
    ```

    Load it in your application:

    <CodeGroup>
```python Python
    from dotenv import load_dotenv
    load_dotenv()
    ```
</CodeGroup>

    <Tip>
    This method keeps your keys organized and makes them available to your Python runtime.
    </Tip>
  </Tab>

  <Tab title="Environment Variables">
    Set environment variables directly in your terminal:

    ```bash
    export OPENAI_API_KEY="your_api_key_here"
    export ANTHROPIC_API_KEY="your_api_key_here"
    ```

    Access them in your application:

    <CodeGroup>
```python Python
    import os
    api_key = os.getenv("OPENAI_API_KEY", "")
    ```
</CodeGroup>
  </Tab>

  <Tab title="System Configuration">
    For production environments, consider using:
    - Docker secrets
    - Kubernetes secrets
    - Cloud provider secret managers (AWS Secrets Manager, etc.)
    - System environment configuration
  </Tab>
</Tabs>

## Agent Parameters

When creating an MCPAgent, you can configure several parameters to customize its behavior:

<CodeGroup>
```python Python
from mcp_use import MCPAgent, MCPClient
from langchain_openai import ChatOpenAI

# Basic configuration
agent = MCPAgent(
    llm=ChatOpenAI(model="gpt-5.5", temperature=0.7),
    client=MCPClient.from_config_file("config.json"),
    max_steps=30
)

# Advanced configuration
agent = MCPAgent(
    llm=ChatOpenAI(model="gpt-5.5", temperature=0.7),
    client=MCPClient.from_config_file("config.json"),
    max_steps=30,
    server_name=None,
    auto_initialize=True,
    memory_enabled=True,
    system_prompt="Custom instructions for the agent",
    additional_instructions="Additional guidelines for specific tasks",
    disallowed_tools=["file_system", "network", "shell"]  # Restrict potentially dangerous tools
)
```
</CodeGroup>

### Available Parameters

- `llm`: Any LangChain-compatible language model (required)
- `client`: The MCPClient instance (optional if connectors are provided)
- `connectors`: List of connectors if not using client (optional)
- `server_name`: Name of the server to use (optional)
- `max_steps`: Maximum number of steps the agent can take (default: 5)
- `auto_initialize`: Whether to initialize automatically (default: False)
- `memory_enabled`: Whether to enable memory (default: True)
- `system_prompt`: Custom system prompt (optional)
- `system_prompt_template`: Custom system prompt template (optional)
- `additional_instructions`: Additional instructions for the agent (optional)
- `disallowed_tools`: List of tool names that should not be available to the agent (optional)
- `use_server_manager`: Enable dynamic server selection (default: False)

## Sending Files with HumanMessage

Pass files directly inside a `HumanMessage` instead of using a separate parameter. The agent accepts either a plain string or a `HumanMessage`, so you can include both text and file content together:

<CodeGroup>
```python Python
import asyncio
import base64
from pathlib import Path

from langchain_core.messages import HumanMessage
from langchain_openai import ChatOpenAI

from mcp_use import MCPAgent, MCPClient


async def main():
    client = MCPClient(config = {})
    agent = MCPAgent(
        llm=ChatOpenAI(model="gpt-5.5", temperature=0),
        client=client,
    )

    image_bytes = base64.b64encode(
        Path("static/logo-gh.jpg").read_bytes()
    ).decode("utf-8")
    human_message = HumanMessage(
        content=[
            {"type": "text", "text": "What is in this image?"},
            {
                "type": "image",
                "base64": image_bytes,
                "mime_type": "image/jpeg",
            },
        ]
    )

    async for step in agent.stream(human_message, max_steps=10):
        print(step)


if __name__ == "__main__":
    asyncio.run(main())
```
</CodeGroup>

<Note>
Remote agents currently accept text-only `HumanMessage` content. Use a local agent when you need to pass files.
</Note>

## Tool Access Control

You can restrict which tools are available to the agent for security or to limit its capabilities. Here's a complete example showing how to set up an agent with restricted tool access:

<CodeGroup>
```python Python
import asyncio
import os
from dotenv import load_dotenv
from langchain_openai import ChatOpenAI
from mcp_use import MCPAgent, MCPClient

async def main():
    # Load environment variables
    load_dotenv()

    # Create configuration dictionary
    config = {
      "mcpServers": {
        "playwright": {
          "command": "npx",
          "args": ["@playwright/mcp@latest"],
          "env": {
            "DISPLAY": ":1"
          }
        }
      }
    }

    # Create MCPClient from configuration dictionary
    client = MCPClient(config)

    # Create LLM
    llm = ChatOpenAI(model="gpt-5.5")

    # Create agent with restricted tools
    agent = MCPAgent(
        llm=llm,
        client=client,
        max_steps=30,
        disallowed_tools=["file_system", "network"]  # Restrict potentially dangerous tools
    )

    # Run the query
    result = await agent.run(
        "Find the best restaurant in San Francisco USING GOOGLE SEARCH",
    )
    print(f"\nResult: {result}")

if __name__ == "__main__":
    asyncio.run(main())
```
</CodeGroup>

You can also manage tool restrictions dynamically:

<CodeGroup>
```python Python
# Update restrictions after initialization
agent.set_disallowed_tools(["file_system", "network", "shell", "database"])
await agent.initialize()  # Reinitialize to apply changes

# Check current restrictions
restricted_tools = agent.get_disallowed_tools()
print(f"Restricted tools: {restricted_tools}")
```
</CodeGroup>

This feature is useful for:

- Restricting access to sensitive operations
- Limiting agent capabilities for specific tasks
- Preventing the agent from using potentially dangerous tools
- Focusing the agent on specific functionality

## Working with Adapters Directly

If you want more control over how tools are created, you can work with the adapters directly. The `BaseAdapter` class provides a unified interface for converting MCP tools to various framework formats, with `LangChainAdapter` being the most commonly used implementation.

```python
import asyncio
from langchain.agents import create_agent
from langchain.chat_models import init_chat_model

from mcp_use.client import MCPClient
from mcp_use.adapters import LangChainAdapter

async def main():
    # Initialize client
    client = MCPClient.from_config_file("browser_mcp.json")

    # Create an adapter instance
    adapter = LangChainAdapter()

    # Get tools directly from the client
    tools = await adapter.create_tools(client)

    # Use the tools with any LangChain agent
    model = init_chat_model("gpt-5.5", temperature=0.5)

    agent = create_agent(
        model=model,
        tools=tools,
        system_prompt="You are a helpful assistant with access to powerful tools.",
    )

    result = await agent.ainvoke(
        {"messages": [{"role": "user", "content": "Search for information about climate change"}]}
    )
    print(result)

if __name__ == "__main__":
    asyncio.run(main())
```

The adapter pattern makes it easy to:

1. Create tools directly from an MCPClient
2. Filter or customize which tools are available
3. Integrate with different agent frameworks

**Benefits of Direct Adapter Usage:**
- **Flexibility**: More control over tool creation and management
- **Custom Integration**: Easier to integrate with existing LangChain workflows
- **Advanced Filtering**: Apply custom logic to tool selection and configuration
- **Framework Agnostic**: Potential for future adapters to other frameworks

## Server Manager

The Server Manager is an agent-level feature that enables dynamic server selection for improved performance with multi-server setups.

### Enabling Server Manager

To improve efficiency and potentially reduce agent confusion when many tools are available, you can enable the Server Manager by setting `use_server_manager=True` when creating the `MCPAgent`.

<CodeGroup>
```python Python
# Enable server manager for automatic server selection
agent = MCPAgent(
    llm=llm,
    client=client,
    use_server_manager=True  # Enable dynamic server selection
)
```
</CodeGroup>

### How It Works

When enabled, the agent will automatically select the appropriate server based on the tool chosen by the LLM for each step. This avoids connecting to unnecessary servers and can improve performance with large numbers of available servers.

<CodeGroup>
```python Python
# Multi-server setup with server manager
client = MCPClient.from_config_file("multi_server_config.json")
agent = MCPAgent(
    llm=llm,
    client=client,
    use_server_manager=True
)

# The agent automatically selects servers based on tool usage
result = await agent.run(
    "Search for a place in Barcelona on Airbnb, then Google nearby restaurants."
)
```
</CodeGroup>

### Benefits

- **Performance**: Only connects to servers when their tools are actually needed
- **Reduced Confusion**: Agents work better with focused tool sets rather than many tools at once
- **Resource Efficiency**: Saves memory and connection overhead
- **Automatic Selection**: No need to manually specify `server_name` for most use cases
- **Scalability**: Better performance with large numbers of servers

### When to Use

- **Multi-server environments**: Essential for setups with 3+ servers
- **Resource-constrained environments**: When memory or connection limits are a concern
- **Complex workflows**: When agents need to dynamically choose between different tool categories
- **Production deployments**: For better resource management and performance

For more details on server manager implementation, see the [Server Manager](./server-manager) guide.

## Memory Configuration

MCPAgent supports conversation memory to maintain context across interactions:

<CodeGroup>
```python Python
# Enable memory (default)
agent = MCPAgent(
    llm=llm,
    client=client,
    memory_enabled=True
)

# Disable memory for stateless interactions
agent = MCPAgent(
    llm=llm,
    client=client,
    memory_enabled=False
)
```
</CodeGroup>

## System Prompt Customization

You can customize the agent's behavior through system prompts:

### Custom System Prompt

<CodeGroup>
```python Python
custom_prompt = """
You are a helpful assistant specialized in data analysis.
Always provide detailed explanations for your reasoning.
When working with data, prioritize accuracy over speed.
"""

agent = MCPAgent(
    llm=llm,
    client=client,
    system_prompt=custom_prompt
)
```
</CodeGroup>

### Additional Instructions

Add task-specific instructions without replacing the base system prompt:

<CodeGroup>
```python Python
agent = MCPAgent(
    llm=llm,
    client=client,
    additional_instructions="Focus on finding recent information from the last 6 months."
)
```
</CodeGroup>

### System Prompt Templates

For more advanced customization, you can provide a custom system prompt template:

```python
from langchain_core.prompts import ChatPromptTemplate

custom_template = ChatPromptTemplate.from_messages([
    ("system", "You are an expert {domain} assistant. {instructions}"),
    ("human", "{input}"),
    # ... other message templates
])

agent = MCPAgent(
    llm=llm,
    client=client,
    system_prompt_template=custom_template
)
```

## Performance Configuration

Configure agent performance characteristics:

```python
# Limit execution steps
agent = MCPAgent(
    llm=llm,
    client=client,
    max_steps=10  # Prevent runaway execution
)

# Enable server manager for better performance with multiple servers
agent = MCPAgent(
    llm=llm,
    client=client,
    use_server_manager=True  # Only connects servers when needed
)

# Limit concurrent servers (if not using server manager)
agent = MCPAgent(
    llm=llm,
    client=client,
    max_concurrent_servers=3
)
```

## Debugging Configuration

Enable debugging features during development:

```python
# Enable verbose logging
agent = MCPAgent(
    llm=llm,
    client=client,
    verbose=True,
    debug=True
)

# Set debug level programmatically
import mcp_use
mcp_use.set_debug(2)  # Full verbose logging
```

## Agent Initialization

Control when and how the agent initializes:

```python
# Auto-initialize on creation
agent = MCPAgent(
    llm=llm,
    client=client,
    auto_initialize=True
)

# Manual initialization for more control
agent = MCPAgent(
    llm=llm,
    client=client,
    auto_initialize=False
)

# Initialize manually when ready
await agent.initialize()
```

## Error Handling

Configure how the agent handles errors:

```python
# Set timeout for agent operations
agent = MCPAgent(
    llm=llm,
    client=client,
    timeout=60  # 60 seconds timeout
)

# Configure retry behavior (if supported by LLM)
llm = ChatOpenAI(
    model="gpt-5.5",
    max_retries=3,
    retry_delay=2
)

agent = MCPAgent(llm=llm, client=client)
```

### Automatic Error Retry

The agent includes automatic error handling that catches tool failures and allows the LLM to retry with corrected input. This is controlled by the `retry_on_error` parameter (default: `True`).

```python
# Default: Error handling enabled
agent = MCPAgent(
    llm=llm,
    client=client,
    retry_on_error=True  # Default
)

# Disable automatic error handling (errors will halt execution)
agent = MCPAgent(
    llm=llm,
    client=client,
    retry_on_error=False
)
```

**How it works:**

When `retry_on_error=True`, the agent uses middleware to intercept tool errors:

1. **Catches all tool errors** including validation errors, connection errors, and execution errors
2. **Formats the error** into a clear, actionable message
3. **Returns error to LLM** as a tool result (not an exception)
4. **LLM reads the error** and can retry with corrected input
5. **Execution continues** based on the LLM's decision

**Example Flow:**

```python
# First attempt - missing required field
# Tool call: create_board_column(board_id="123", body={"name": "Assignee"})
# Error returned: "Tool input validation failed. Please fix the following errors and retry:
#   - body.metaData: Field required (type: missing)"

# Second attempt - LLM automatically retries with correction
# Tool call: create_board_column(board_id="123", body={"name": "Assignee", "metaData": {}})
# Success: Column created
```

<Note>
The automatic retry mechanism works for all error types and is handled by the middleware, ensuring errors are caught at the right level - including Pydantic validation errors that occur before tool execution.
</Note>

**Error Message Formatting:**

Different error types receive different formatting:

- **Validation Errors**: Field-by-field breakdown showing exactly what's wrong
  ```
  Tool input validation failed. Please fix the following errors and retry:
    - body.metaData: Field required (type: missing)
    - timeout: Input should be a valid integer (type: int_type)
  ```

- **Other Errors**: Error type and description
  ```
  ConnectionError: Failed to connect to MCP server
  ```

**When to disable `retry_on_error`:**

Set `retry_on_error=False` when:
- Debugging tool issues and you want immediate failure
- You want strict validation without retry attempts
- Running in environments where retries could cause issues

## Best Practices

1. **LLM Selection**: Use models with tool calling capabilities
2. **Step Limits**: Set reasonable `max_steps` to prevent runaway execution
3. **Tool Restrictions**: Use `disallowed_tools` for security
4. **Memory Management**: Disable memory for stateless use cases
5. **Server Manager**: Enable for multi-server setups
6. **System Prompts**: Customize for domain-specific tasks
7. **Error Handling**: Implement proper timeout and retry logic
8. **Testing**: Test agent configurations in development environments

## Common Issues

1. **No Tools Available**: Check client configuration and server connections
2. **Tool Execution Failures**: Enable verbose logging and check tool arguments
3. **Memory Issues**: Disable memory or limit concurrent servers
4. **Timeout Errors**: Increase `max_steps` or agent timeout values

For detailed troubleshooting, see the [Security](/python/development/security) guide.