Built-in Tools

In CrewAI, tools extend what an agent can do beyond raw LLM reasoning. They are callable capabilities—search APIs, file readers, scrapers, interpreters—that let agents gather facts, read documents, and interact with external systems instead of only generating text from training data.

Assigning Tools to an Agent

Pass a list of tool instances to the agent’s tools parameter. The LLM can invoke these tools when a task requires real data or side effects.

from crewai import Agent
from crewai_tools import SerperDevTool
 
researcher = Agent(
    role="Research Analyst",
    goal="Find accurate, up-to-date information on the web",
    backstory="You are meticulous about citing sources and verifying facts.",
    tools=[SerperDevTool()],
    verbose=True,
)

Assigning Tools to a Task

You can also set tools on a Task. For that task only, the task’s tool list overrides the agent’s default tools. Use this when one step needs search but another should stay “LLM-only,” or when different tasks need different integrations.

from crewai import Task
 
task_search = Task(
    description="Search the web for the latest CrewAI release notes.",
    expected_output="A short summary with links or titles found.",
    agent=researcher,
    tools=[SerperDevTool()],
)
 
task_summarize = Task(
    description="Summarize the previous task’s output in two sentences.",
    expected_output="Two-sentence summary.",
    agent=researcher,
    tools=[],  # no tools for this task — overrides agent tools
)

Major Built-in Tools (crewai_tools)

These ship in the crewai-tools package and cover common research and document workflows.

Install extras as needed; many tools depend on API keys or local paths you configure in the environment.

Example: Researcher with SerperDevTool

Set your Serper API key (and model provider key) before running:

import os
 
os.environ["SERPER_API_KEY"] = "your-serper-api-key"
# e.g. OPENAI_API_KEY for the default LLM

Full crew sketch:

from crewai import Agent, Crew, Process, Task
from crewai_tools import SerperDevTool
 
search_tool = SerperDevTool()
 
researcher = Agent(
    role="Web Researcher",
    goal="Answer questions using current web results",
    backstory="You search first, then synthesize; you never invent URLs.",
    tools=[search_tool],
    verbose=True,
)
 
task = Task(
    description=(
        "Search for what CrewAI is and list three bullet points: "
        "what it is, one main use case, and the official site or docs URL if found."
    ),
    expected_output="Three markdown bullets with factual, search-backed points.",
    agent=researcher,
    tools=[search_tool],
)
 
crew = Crew(
    agents=[researcher],
    tasks=[task],
    process=Process.sequential,
    verbose=True,
)
 
result = crew.kickoff()
print(result)

Tool Errors and Caching

• Errors: Tools call real APIs and filesystems. Network failures, rate limits, missing keys, or invalid paths surface as tool errors. The agent may retry or adjust depending on your LLM and crew settings; design tasks so failures produce a clear expected_output (e.g. “state that search failed and why”).
• Caching: On Agent, set cache=True to cache tool results for identical invocations within a run, which can reduce duplicate API calls and cost. Use it when repeated reads or searches would otherwise hit the same inputs.

efficient_agent = Agent(
    role="Analyst",
    goal="Analyze without redundant tool calls",
    backstory="You reuse prior tool output when the question repeats.",
    tools=[SerperDevTool()],
    cache=True,
    verbose=True,
)

Key Takeaways

• Tools bridge agents to external systems (search, files, code execution, documents).
• tools=[...] on Agent defines the default toolkit for that agent.
• tools on Task overrides the agent’s tools for that task only.
• crewai_tools provides many ready-made tools; pair each with the right API keys and paths.
• cache=True on an agent can cut duplicate tool work; handle tool errors explicitly in task descriptions and expected outputs.