Use Built-in Tools

This guide shows how to give a Custodian Agent access to the web and to documents it was not trained on, using the built-in tools web_search, scrape_url, and parse_document.

For the full reference, see Built-in Tools for Python or TypeScript.

When to use a tool

Situation	Approach
The agent should answer from your own stable documents	Data source / RAG
The agent needs current information from the web	web_search
The agent should read one specific page you name	scrape_url
The agent should extract text from a document by URL, one time	parse_document

Give an agent web search

Python

from custodian_labs import Custodian

app = Custodian(
    model="gpt-4o",
    system_prompt="Answer using current web results when they help. Cite your sources.",
    tools=["web_search", "scrape_url"],
).deploy()

reply = app.chat("What changed in the latest OpenAI API release?")
print(reply.response)

TypeScript

import { Custodian } from "@custodianlabs/sdk";

const app = await new Custodian({
  model: "gpt-4o",
  systemPrompt: "Answer using current web results when they help. Cite your sources.",
  tools: ["web_search", "scrape_url"],
}).deploy();

const reply = await app.chat("What changed in the latest OpenAI API release?");
console.log(reply?.response);

Combine tools with a workflow team

A common pattern: one agent researches with web_search, the next writes from its notes. In workflow mode the agents run in a fixed order and each output feeds the next.

Python

from custodian_labs import Custodian, CustodianSquad

squad = CustodianSquad(
    custodians=[
        Custodian(
            name="research",
            model="gpt-4o",
            system_prompt="Research the topic. Produce notes with sources.",
            tools=["web_search"],
            output_key="research_notes",
        ),
        Custodian(
            name="writer",
            model="gpt-4o",
            system_prompt="Write a one-paragraph brief from the research notes.",
        ),
    ],
    routing_mode="workflow",
    workflow_order=["research", "writer"],
)
app = squad.deploy()
reply = app.chat("Give me a brief on RAG evaluation methods.")
print(reply.response)

TypeScript

import { Agent, AgentTeam } from "@custodianlabs/sdk";

const team = new AgentTeam({
  agents: [
    new Agent({
      name: "research",
      model: "gpt-4o",
      systemPrompt: "Research the topic. Produce notes with sources.",
      tools: ["web_search"],
      outputKey: "research_notes",
    }),
    new Agent({
      name: "writer",
      model: "gpt-4o",
      systemPrompt: "Write a one-paragraph brief from the research notes.",
    }),
  ],
  routingMode: "workflow",
  workflowOrder: ["research", "writer"],
});

const app = await team.deploy();
const reply = await app.chat("Give me a brief on RAG evaluation methods.");
console.log(reply?.response);

Cost and latency

Each tool call is an extra round trip. Tell the agent in its system prompt when to use a tool ("search only when the answer depends on recent events") rather than letting it reach for one on every message.

External connections

Gmail (read / send / reply from an agent) is coming soon.

Next steps

• Python → Built-in Tools
• TypeScript → Built-in Tools
• Build a RAG for sensitive data