AutoGPT Integration Guide

Integrate AgentsMon with AutoGPT to monitor autonomous agent commands, web browsing, and file operations.

Installation

``bash


pip install agentsmon

This installs the AgentsMon Python SDK which includes AgentsMonAutoGPTHooks for AutoGPT integration.



SDK Integration

The Python SDK provides a ready-made hooks class that monitors AutoGPT's command execution, browsing, file operations, and thinking phases. No need to write your own plugin -- just import and use:

`python


from agentsmon.autogpt import AgentsMonAutoGPTHooks
import os

hooks = AgentsMonAutoGPTHooks(
    endpoint=os.getenv("AGENTSMON_URL", "http://localhost:18800"),
    agent_id=os.getenv("AUTOGPT_AGENT_ID", "autogpt-main")
)

The AgentsMonAutoGPTHooks class provides:

on_command(command_name, arguments) -- tracks command execution

on_thinking(thoughts) -- tracks the thinking/planning phase

on_browse(url) -- tracks web browsing

on_write_file(filename, content_length) -- tracks file writes

on_read_file(filename) -- tracks file reads



All events are sent asynchronously and non-blocking -- if AgentsMon is unreachable, your agent continues unaffected.

Integration with AutoGPT

Option 1: Direct Hook (Recommended)

Add to your AutoGPT main.py or agent loop:

`python


from agentsmon.autogpt import AgentsMonAutoGPTHooks
import os

hooks = AgentsMonAutoGPTHooks(
    endpoint=os.getenv("AGENTSMON_URL", "http://localhost:18800"),
    agent_id=os.getenv("AUTOGPT_AGENT_ID", "autogpt-main")
)

In your command execution loop:
def execute_command(command_name: str, arguments: dict) -> str:
    # Monitor the command
    hooks.on_command(command_name, arguments)

    # Handle specific commands
    if command_name == "browse_website":
        hooks.on_browse(arguments.get("url", ""))
    elif command_name == "write_to_file":
        hooks.on_write_file(arguments.get("filename", ""))
    elif command_name == "read_file":
        hooks.on_read_file(arguments.get("filename", ""))

    # Execute the actual command
    return original_execute(command_name, arguments)



Option 2: Webhook Forwarder

If you cannot modify AutoGPT source, use the SDK's log forwarder:

`python


from agentsmon.autogpt import AgentsMonAutoGPTHooks

hooks = AgentsMonAutoGPTHooks(
    endpoint="http://localhost:18800",
    agent_id="autogpt-agent"
)

The SDK can also tail AutoGPT logs and forward events automatically:
See agentsmon.autogpt.LogForwarder for details



Shield Mode

AutoGPT is autonomous -- it can execute any command. AgentsMon Shield mode lets you block unsafe actions before they run:

`python


from agentsmon import AgentsMonGuard
from agentsmon.autogpt import AgentsMonAutoGPTHooks
import os

guard = AgentsMonGuard(
    endpoint=os.getenv("AGENTSMON_URL", "http://localhost:18800"),
    agent_id="autogpt-main"
)

hooks = AgentsMonAutoGPTHooks(
    endpoint=os.getenv("AGENTSMON_URL", "http://localhost:18800"),
    agent_id="autogpt-main"
)

def execute_command(command_name: str, arguments: dict) -> str:
    # Check with guard before executing
    result = guard.check_command(f"{command_name} {arguments}")
    if result.blocked:
        return f"BLOCKED by AgentsMon: {result.reason}"

    # Also check file access
    if command_name in ("write_to_file", "read_file"):
        file_result = guard.check_file(arguments.get("filename", ""), command_name.split("_")[0])
        if file_result.blocked:
            return f"BLOCKED by AgentsMon: {file_result.reason}"

    # Monitor the command
    hooks.on_command(command_name, arguments)

    # Execute the actual command
    return original_execute(command_name, arguments)



Shield mode calls AgentsMon's security engines (sandbox monitor, behavioral analyzer, prompt injection scanner) synchronously and returns a verdict before the action runs.

Event Mapping

| AutoGPT Action | AgentsMon Event | Data Captured |
|---------------|-----------------|---------------|

| Command execution | command | Command name, arguments |

| Thinking/planning | command | Thought summary |

| Web browsing | http_request | URL, method |

| File write | file_access | File path, write operation |

| File read | file_access | File path, read operation |



Verification

`bash


Check events are flowing
curl http://localhost:18800/api/events?platform=autogpt

Check agent was registered
curl http://localhost:18800/api/agents?platform=autogpt

Check for dangerous commands
curl http://localhost:18800/api/events/anomalies?platform=autogpt



Docker Deployment

`yaml


services:
  agentsmon:
    build: ./agentsmon/backend
    ports: ["18800:18800"]

  autogpt:
    image: significantgravitas/auto-gpt
    environment:
      - AGENTSMON_URL=http://agentsmon:18800
    depends_on: [agentsmon]



In your AutoGPT Dockerfile or requirements:


agentsmon



Security Considerations

AutoGPT is autonomous -- it can execute any command. AgentsMon is especially important here:

Sandbox monitoring: Detects container escape attempts (/api/sandbox/analyze-command)


Network isolation: Alerts on suspicious outbound connections
File access tracking: Monitors sensitive file access patterns
Trust scoring: Starts AutoGPT at trust level 0 (Intern) with read-only permissions
Shield mode: Blocks dangerous commands before execution (see above)

`bash


Check trust level for your AutoGPT agent
curl http://localhost:18800/api/trust/level/autogpt-main

View sandbox violations
curl http://localhost:18800/api/sandbox/violations?agentId=autogpt-main