Web access

Two default capabilities put the web within the agent’s reach: web_fetch reads a specific page, and web_search looks things up. Both are on by default.

web_fetch

web_fetch performs an HTTP GET on a URL, reduces the response to text (stripping HTML to plain content, capped so a huge page does not overwhelm the model), and asks a small, fast summarizer model to answer your prompt over that content. It returns the answer, not the raw page, so a long article never floods the conversation. Fetched content is cached for 15 minutes, so repeated fetches of the same URL are fast while each prompt is answered fresh.

The tool takes a url and an optional prompt. With a prompt, the summarizer answers over the page; leave the prompt empty and it returns the page text verbatim instead. Only http and https URLs are allowed.

The summarizer model

By default the summarizer runs on the small, fast model for the main agent’s provider, so its credentials line up with the main model (on Amazon Bedrock, the summarizer follows the main model’s family, Anthropic-on-Bedrock or OpenAI-on-Bedrock). Override it with builtin_tools={"web_fetch": {"model": ...}}, which takes the same forms as model:

Strands harness

Python

from strands_harness import create_harness

agent = create_harness(builtin_tools={"web_fetch": {"model": "bedrock/global.anthropic.claude-haiku-4-5-20251001-v1:0"}})

TypeScript

import { createHarness } from '@strands-agents/harness'

const agent = await createHarness({
  builtinTools: { web_fetch: { model: 'bedrock/global.anthropic.claude-haiku-4-5-20251001-v1:0' } },
})

SDK

Python

# pip install strands-agents
from strands import Agent
from strands.vended_tools import web_fetch

# No summarizer-model override in the Strands Harness SDK; web_fetch returns page text
agent = Agent(tools=[web_fetch])

TypeScript

// npm install @strands-agents/sdk
import { Agent } from '@strands-agents/sdk'
import { httpRequest } from '@strands-agents/sdk/vended-tools/http-request'

// No web_fetch or summarizer model; httpRequest returns raw HTTP
const agent = new Agent({ tools: [httpRequest] })

web_search

Where the model provider has its own search, web_search is a switch, not a tool object. It turns on the model provider’s own native web search so the agent can look things up mid-answer. Native search exists on OpenAI, Anthropic (Python library), Google, and the GPT-5 and GPT-6 models on bedrock-mantle, so web_search takes effect there.

On providers without native search (Amazon Bedrock Converse, other bedrock-mantle models, and Model instances), web_search is a no-op when it is on only by default: Strands harness logs a warning and builds the agent without it. Requesting it explicitly on such a provider (by listing web_search in your own builtin_tools) fails at construction instead, so a deliberate request never silently does nothing.

To search on those providers anyway, set {"web_search": "exa"} in builtin_tools. The harness then builds a web_search tool backed by Exa’s hosted search instead of relying on the provider. It needs no setup: the keyless free tier covers getting started. Exa is a third party, so every query the agent makes is sent to Exa and handled under its terms; weigh that before turning it on for sensitive work. Set EXA_API_KEY in the environment to lift the free tier’s rate limit.

Selecting them

Both are selected by name through builtin_tools. Drop web_fetch or web_search from the list to turn either off; see add tools and instructions for how selection works. To change the summarizer model, see choose a model and reasoning.