Backend Tools Agent with CrewAI

Let a CrewAI agent call backend APIs (CRM, weather, ticketing) and stream results to CometChat via NDJSON.

Put your APIs behind controlled tools so the agent can fetch data or perform actions while CometChat streams results live.

---

What you’ll build

• A CrewAI agent configured with one or more backend tools.
• Tool functions that validate inputs, call external APIs, and raise exceptions on failure.
• A /kickoff endpoint that surfaces tool call events to CometChat.

---

Prerequisites

• CrewAI project from crew-ai.mdx
• API keys for any services you call (store them in .env)

---

Steps

1. Model your tools

   Keep each tool focused (single responsibility) and add clear docstrings so the LLM knows when to call them.

2. Validate inputs

   Check for required fields and raise exceptions for bad data—don’t return error strings.

3. Register in CrewAI

   Add tools to the agent in crew.py and set process=Process.sequential or another flow as needed.

4. Stream via FastAPI

   Reuse the NDJSON stream from crew-ai.mdx so CometChat can render tool progress.

---

Sample tool: fetch deals from a CRM

src/crew_demo/tools/get_deals.py

import os, httpx, json
from crewai.tools import tool


@tool("get_recent_deals")
def get_recent_deals(limit: int = 5) -> str:
    """Fetch recent deals from the CRM."""
    api_key = os.getenv("CRM_API_KEY")
    base_url = os.getenv("CRM_BASE_URL", "https://api.example-crm.com")
    if not api_key:
        raise Exception("CRM_API_KEY not set")

    resp = httpx.get(f"{base_url}/deals", params={"limit": limit}, headers={"Authorization": f"Bearer {api_key}"}, timeout=10.0)
    resp.raise_for_status()
    deals = resp.json().get("deals", [])
    return json.dumps({"deals": deals[:limit]})

Register in crew.py:

from crew_demo.tools.get_deals import get_recent_deals

@agent
def backend(self) -> Agent:
    return Agent(
        config=self.agents_config["backend"],  # type: ignore[index]
        tools=[get_recent_deals],
        verbose=False,
        memory=False,
    )

---

Agent/task configuration

src/crew_demo/config/agents.yaml

backend:
  role: Backend Integrations Agent
  goal: Call approved APIs and summarize results
  backstory: >
    Always validate inputs, call the correct tool, and summarize results clearly.
    Never expose raw errors—raise exceptions and let the server handle them.

src/crew_demo/config/tasks.yaml

backend_task:
  description: >
    Use the appropriate tool to answer: {user_message}
  expected_output: >
    A short summary plus any key fields (totals, status, links).
  agent: backend

---

Connect to CometChat

• Provider: CrewAI
• Agent ID: backend
• Deployment URL: your public /kickoff
• Optional headers: { "Authorization": "Bearer <token>" }

Ensure your FastAPI service returns NDJSON with tool events so users can see when an API call is running.

---