Skip to main content

​
Hooking AutoGPT into TaskGuru

Autonomous agents like AutoGPT, BabyAGI, and Devin are incredible at breaking down massive goals into smaller steps. However, their internal memory is often a black box CLI terminal, making it difficult for humans to understand what the agent is actually planning to do next. TaskGuru solves this. By writing a simple custom plugin for AutoGPT, you can force the agent to use TaskGuru as its “Workspace”.

​
Use Case: AI Software Developer

Imagine you give AutoGPT the prompt: “Build a weather web app in React.” Instead of AutoGPT just vomiting terminal text, we teach it to:
  1. Create a Board in TaskGuru called ”⛅ Weather App Build”.
  2. Generate all the sub-tasks it thinks it needs (e.g., “Setup Vite”, “Fetch Weather API”, “Design CSS”).
  3. Assign the tricky tasks (like “Approve CSS Design”) to you, the human.
  4. Mark tasks as COMPLETED as it finishes writing the code.
You can literally watch the AI move cards across the Kanban board.

​
Building the AutoGPT Custom Tool (Python)

If you are using the modern AutoGPT architecture or any generic LangChain agent toolkit, you can wrap the TaskGuru REST API directly into a custom tool.

​
Step 1: Wrap the Create Task Endpoint

import os
import requests
from typing import Optional, Type
from pydantic import BaseModel, Field
from langchain.tools import BaseTool

class CreateTaskguruTaskInput(BaseModel):
    name: str = Field(description="The name/title of the task the AI wants to complete.")
    description: str = Field(description="Markdown description of the steps needed.")
    board_id: str = Field(description="The UUID of the TaskGuru board to place this task on.")
    urgency: Optional[str] = Field("medium", description="none, low, medium, high, or urgent")

class CreateTaskguruTaskTool(BaseTool):
    name = "create_taskguru_task"
    description = "Use this tool to break down a large goal into smaller actionable tasks on the human's TaskGuru board."
    args_schema: Type[BaseModel] = CreateTaskguruTaskInput

    def _run(self, name: str, description: str, board_id: str, urgency: str = "medium") -> str:
        headers = {
            "Authorization": f"Bearer {os.getenv('TASKGURU_API_KEY')}",
            "Content-Type": "application/json"
        }

        payload = {
            "tasks": [{
                "name": name,
                "description": description,
                "board_id": board_id,
                "urgency": urgency
            }]
        }

        # Notice we are calling the precise OpenAPI spec endpoint
        response = requests.post(
            "https://api.taskguru.com/v1/tasks",
            json=payload,
            headers=headers
        )

        if response.status_code == 201:
            return f"Successfully created task '{name}'. Human can now view it."
        else:
            return f"Failed to create task: {response.text}"

    def _arun(self, name: str, description: str, board_id: str):
        raise NotImplementedError("create_taskguru_task does not support async yet")

​
Step 2: Providing Agent Scratchpad Context

AutoGPT frequently forgets its previous steps when the context window fills up. TaskGuru’s threading architecture solves this. Before AutoGPT attempts a Python script, teach it to pull previous context from the Task’s message thread: 
def read_task_context(task_id: str) -> str:
    """Reads all messages attached to a task to regain context."""
    headers = {"Authorization": f"Bearer {os.getenv('TASKGURU_API_KEY')}"}
    response = requests.get(f"https://api.taskguru.com/v1/tasks/{task_id}/messages", headers=headers)

    messages = response.json()
    context = ""
    for msg in messages:
        context += f"Previous note: {msg['content']}\n"

    return context

​
Why this is better than local JSON files

By forcing AutoGPT to use TaskGuru instead of its local workspace/ directory for task tracking:
  • You get a beautiful mobile app to monitor your AI engine.
  • You can step in and leave a comment on a task: “Hey AutoGPT, don’t use React for this, use vanilla JS.” The agent will read that message on the next loop and adjust its plan.