Build with Browser Use - SambaNova Integration Guide

Browser Use is an open-source project that empowers AI agents to control web browsers, enabling tasks such as automated web navigation and data extraction.

Prerequisites

Before you begin, ensure you have:

Python 3.11 or later. Browser Use declares requires-python >=3.11, so an older interpreter fails to resolve the package rather than reporting a clear error
A SambaCloud API key. Get one by creating a SambaCloud account and generating a key from the API keys tab

Browser Use needs two things from a model, and they are separate capabilities. It sends every screenshot to the model while use_vision=True (the default), so the model must accept image input. It also asks for each next action as JSON using response_format with type: json_schema, so the model must support structured output. It does not use tool calling, so a model’s absence from the function calling list does not rule it out.This guide uses gemma-4-31B-it, which accepts image input and has constrained decoding enabled, so json_schema is enforced as the response is generated. See Choosing a model for the text-only alternative.

1. Install Browser Use

create environment

pip install uv
uv venv --python 3.12

activate environment

source .venv/bin/activate
# On Windows use `.venv\Scripts\activate`

install browser-use and chromium

uv pip install "browser-use>=0.13,<0.14"
browser-use install

The second command is a separate step and easy to skip. Browser Use drives its own Chromium build, which uv pip install does not fetch, so without it the first run fails when it tries to launch a browser. Expect about 180 MB and a few minutes; it prints Installation complete. when it finishes.

Run browser-use install inside the activated virtual environment, not as uvx browser-use install. uvx ignores your environment and resolves its own copy of Browser Use, so you can end up with a browser installed for a different version than the one your code imports.

2. Configure your SambaCloud key

Create a .env file in your project directory:

.env

touch .env

On Windows, use echo. > .env

Add your SambaCloud API key:

.env

SAMBANOVA_API_KEY=your-sambanova-api-key

3. Run your first agent

Browser Use has no dedicated SambaNova class. Connect to SambaCloud through ChatOpenAI with a custom base_url – the same pattern Browser Use uses for any OpenAI-compatible provider.

agent.py

from browser_use import Agent, ChatOpenAI
from dotenv import load_dotenv
import asyncio
import os
 
load_dotenv()
 
llm = ChatOpenAI(
    model="gemma-4-31B-it",
    base_url="https://api.sambanova.ai/v1",
    api_key=os.getenv("SAMBANOVA_API_KEY"),
)
 
async def main():
    task = "Find the number 1 post on Show HN"
    agent = Agent(task=task, llm=llm)
    result = await agent.run(max_steps=25)
    print(result.final_result())
 
if __name__ == "__main__":
    asyncio.run(main())

Run it:

python agent.py

Browser Use opens a Chromium window, logs each step it takes to the terminal, and closes when the task finishes. The step log comes from Browser Use itself; the answer to your task is the value final_result() returns, which is why the example captures result rather than discarding it.

Always pass max_steps. It defaults to 500, and every step on a vision model sends another screenshot, so a task that never converges keeps calling the model until it hits that ceiling. Start low, around 25, and raise it only for tasks that genuinely need more navigation.

Keep credentials out of the prompt

Pass anything secret through sensitive_data rather than writing it into the task, and scope it to the domains that are allowed to receive it. Browser Use substitutes the value when it fills a field, so the literal never reaches the model:

from browser_use import BrowserProfile

agent = Agent(
    task="Log in to the dashboard and download this month's invoice",
    llm=llm,
    sensitive_data={"https://*.example.com": {"x_password": os.getenv("PORTAL_PASSWORD")}},
    browser_profile=BrowserProfile(allowed_domains=["https://*.example.com"]),
)

Set allowed_domains whenever you use sensitive_data. Without it Browser Use logs Agent(sensitive_data=••••••••) was provided but Browser(allowed_domains=[...]) is not locked down! and warns that a prompt-injection attack on any page the agent visits could expose your credentials. An agent that browses freely with secrets loaded is the main security risk in this integration.

The browser runs visible by default, which is useful while you are building. For an unattended run, set headless=True on the profile:

agent = Agent(task=task, llm=llm, browser_profile=BrowserProfile(headless=True))

Choosing a model

SambaCloud model support for Browser Use depends on the task:

Task	Recommended model	Notes
General browsing with screenshots	`gemma-4-31B-it`	Accepts image input, so it works with the default `use_vision=True`. Constrained decoding is enabled, so the `json_schema` response format is enforced during generation
Text-only flows	`Meta-Llama-3.3-70B-Instruct`	Text only, so you must set `use_vision=False` on the `Agent`. Leaving vision on sends images to a model that cannot read them

Set use_vision=False like this:

agent = Agent(task=task, llm=llm, use_vision=False)

See SambaCloud models for the current catalog and Model deprecations before you reuse a model ID from an older guide, since IDs are retired over time.

Troubleshooting

AuthenticationError: Incorrect API key

Confirm SAMBANOVA_API_KEY is set in .env and that load_dotenv() runs before you construct ChatOpenAI. A key copied with a trailing space or newline also triggers this.

Agent loops without taking useful actions, or errors on image input

Two different causes. If it errors on image input, the model is text only: switch to gemma-4-31B-it, or keep your model and set use_vision=False. If it runs but never picks a useful action, the model is likely not honoring the json_schema response format that Browser Use uses to request each action, so choose a model with structured output support.

NotFoundError: model not found

The model string must match a current SambaCloud model ID exactly, and IDs are case-sensitive and take no provider prefix. Confirm it against SambaCloud models, or list the catalog directly with curl https://api.sambanova.ai/v1/models, which needs no API key. A model that has been retired returns this error even though older guides still reference it.

Chromium fails to launch

Run browser-use install inside the activated virtual environment. Browser Use drives its own Chromium build, which uv pip install browser-use does not fetch, and the command prints Installation complete. when it succeeds. If you ran it as uvx browser-use install, run it again inside the environment: uvx resolves its own copy of Browser Use, so the browser it installed may belong to a different version than the one your code imports.

Next steps

Quickstart – make a plain SambaCloud API call, useful for confirming your key works before adding a browser
SambaCloud models – current model IDs, context lengths, and which accept image input
Structured output – the response_format mechanism Browser Use relies on for each action
API reference – endpoints and parameters
Browser Use documentation – advanced configuration, custom browsers, and production deployment
Browser Use supported models reference – full list of native and OpenAI-compatible integrations

Overview

Developer tools

Frameworks

Interfaces and low-code

Data and infrastructure

Multimodal

Evaluation and monitoring

Browser Use integration guide

Prerequisites

1. Install Browser Use

2. Configure your SambaCloud key

3. Run your first agent

Keep credentials out of the prompt

Choosing a model

Troubleshooting

Next steps

​Prerequisites

​1. Install Browser Use

​2. Configure your SambaCloud key

​3. Run your first agent

​Keep credentials out of the prompt

​Choosing a model

​Troubleshooting

​Next steps

Prerequisites

1. Install Browser Use

2. Configure your SambaCloud key

3. Run your first agent

Keep credentials out of the prompt

Choosing a model

Troubleshooting

Next steps