How to Add ATXP to CrewAI

CrewAI makes it easy to build multi-agent workflows where each agent has a specialized role. ATXP gives each agent real-world capabilities — web access, email, payments, image generation — so the crew can execute tasks beyond text generation.

---

The short answer

from crewai import Agent, Task, Crew
from atxp import AtxpToolkit

tools = AtxpToolkit.from_env().get_tools()

researcher = Agent(role="Researcher", tools=tools, ...)
writer = Agent(role="Writer", tools=tools, ...)
crew = Crew(agents=[researcher, writer], tasks=[...])
crew.kickoff()

AtxpToolkit.from_env().get_tools() returns all ATXP tools as CrewAI-compatible tools. Pass the full list or a subset to each agent based on its role.

---

Prerequisites

pip install crewai atxp langchain-anthropic

npx atxp  # Provisions agent account
export ATXP_API_KEY="your-key-from-npx-atxp"

---

Complete setup: three-agent research crew

from crewai import Agent, Task, Crew, Process
from langchain_anthropic import ChatAnthropic
from atxp import AtxpToolkit

llm = ChatAnthropic(model="claude-3-7-sonnet-latest")
toolkit = AtxpToolkit.from_env()

# Researcher: web access only
researcher = Agent(
    role="Market Researcher",
    goal="Find accurate, current information on AI agent frameworks",
    backstory="Expert at navigating web sources and synthesizing research",
    llm=llm,
    tools=toolkit.get_tools(["web_search", "web_browse"])
)

# Writer: generation and storage
writer = Agent(
    role="Technical Writer",
    goal="Produce clear, accurate technical comparisons",
    backstory="Translates complex technical findings into readable reports",
    llm=llm,
    tools=toolkit.get_tools(["image_generate", "file_store"])
)

# Publisher: communication
publisher = Agent(
    role="Publisher",
    goal="Deliver the completed report to the right recipients",
    backstory="Handles final delivery and archival of research outputs",
    llm=llm,
    tools=toolkit.get_tools(["email_send", "file_store"])
)

# Tasks
research_task = Task(
    description="Research the top 5 AI agent frameworks released in Q1 2026. Find their key capabilities, pricing, and ecosystem support.",
    agent=researcher,
    expected_output="Structured comparison of 5 frameworks with capabilities, pricing, and ecosystem notes"
)

write_task = Task(
    description="Write a technical comparison report based on the research. Generate a summary diagram.",
    agent=writer,
    expected_output="Written report + architecture diagram image URL",
    context=[research_task]
)

publish_task = Task(
    description="Email the completed report to team@company.com and archive to file storage.",
    agent=publisher,
    expected_output="Email confirmation and file storage path",
    context=[write_task]
)

crew = Crew(
    agents=[researcher, writer, publisher],
    tasks=[research_task, write_task, publish_task],
    process=Process.sequential
)

result = crew.kickoff()

---

Tool allocation by role

Definition — Multi-Agent Tool Isolation

Multi-agent tool isolation is the practice of giving each agent in a crew only the specific tools required for its role — a researcher gets web search and browsing, a publisher gets email and file storage, a buyer gets payment tools only. This limits blast radius per agent, keeps each agent's behavior focused on its designated task, and ensures one agent's misbehavior cannot leverage another role's permissions.

Giving each agent only the tools it needs limits blast radius and keeps each agent focused:

Role	Recommended tools	Why
Researcher	web_search, web_browse	Information gathering only
Analyst	web_search, code_execute, file_store	Computation and storage
Writer	image_generate, file_store	Content creation
Publisher	email_send, file_store	Delivery and archival
Buyer	web_browse, payment_make	Commerce only
Orchestrator	Full toolkit	Coordinates, delegates

# Get specific tools by name
researcher_tools = toolkit.get_tools(["web_search", "web_browse"])
buyer_tools = toolkit.get_tools(["web_browse", "payment_make"])
full_tools = toolkit.get_tools()  # All tools

---

Per-agent spending isolation

For multi-crew setups where you want per-agent cost tracking:

# Create named agent accounts
npx atxp --agent "researcher"
npx atxp --agent "writer"
npx atxp --agent "publisher"

# Fund each separately
npx atxp fund --agent "researcher" --amount 20
npx atxp fund --agent "writer" --amount 30

Then load each agent’s account via environment variable:

import os

os.environ["ATXP_API_KEY"] = os.environ["ATXP_RESEARCHER_KEY"]
researcher_tools = AtxpToolkit.from_env().get_tools(["web_search", "web_browse"])

os.environ["ATXP_API_KEY"] = os.environ["ATXP_WRITER_KEY"]
writer_tools = AtxpToolkit.from_env().get_tools(["image_generate", "file_store"])

Each agent’s tool calls draw from its own IOU balance and appear in its own audit log.

---

pip install crewai atxp && npx atxp

10 free IOU tokens on registration. Full docs →

For other frameworks: LangChain → · OpenAI Agents SDK → · Claude Code →

---

Frequently asked questions

How do I add ATXP to CrewAI?

tools=AtxpToolkit.from_env().get_tools() in your Agent definition. Use get_tools(["tool1", "tool2"]) for a subset.

What does ATXP add to CrewAI?

14+ tools as native CrewAI tools — web search, browsing, image gen, email, payments, code execution, file storage.

Can different agents have different tools?

Yes — pass toolkit.get_tools(["web_search", "web_browse"]) to researcher, toolkit.get_tools(["email_send"]) to publisher.

Does each agent need its own ATXP account?

Optional. One account works for a crew. Separate accounts per agent enable per-agent spending isolation and audit trails.

How does payment work?

IOU balance charged per tool call automatically. Agent calls the tool; ATXP handles billing. How ATXP’s IOU model works →

Does it work with CrewAI task delegation?

Yes — tool access follows the agent, not the task. Delegated tasks use the receiving agent’s tool list.