Software Sushi publisher · customer guide

Software Sushi Browser Agent

Ubuntu 24.04 · Browser Use 0.10.1 · Headless Chromium

Run the historical official open-source Browser Use Python agent with a local Chromium browser and your own compatible model provider. This package includes a command-line launcher and a synthetic records-and-form walkthrough. There is no web dashboard or public task API.

Software charge: USD 0.03 per running instance-hour. EC2, EBS, data transfer and model-provider usage are billed separately. Model credits, Browser Use Cloud, hosted browsers and proxy subscriptions are not included.

1. Launch and connect

1. Choose t3.medium or larger, at least 24 GiB of root storage, encrypted gp3, and IMDSv2. No application ingress rule is needed.
2. Connect through AWS Systems Manager Session Manager using your instance IAM configuration. SSH is an alternative when you configure your own key and restricted administrative access.
3. Wait for automatic cloud initialization. Run the included walkthrough as the customer account:

sudo -iu ubuntu bash /opt/sushi/docs/demo.sh

A successful walkthrough prints SUSHI_CUSTOMER_DEMO_PASSED and the actual result directory. Runs are retained beneath /home/ubuntu/browser-use-results. The expected three records are 5, 17 and 20, totaling 42.

2. Connect your model provider

Choose an OpenAI-compatible chat-completions provider and exact model supporting JSON schema structured output. Compatibility depends on that provider and model. Switch to ubuntu and run the interactive setup:

sudo -iu ubuntu
sushi-browser-use configure

Enter the provider base URL, exact model ID and key when prompted. Key entry is hidden; setup sends no model request. Settings are owned by ubuntu, mode 0600, under ~/.config/sushi/browser-use/provider.json. Existing settings are preserved; setup refuses overwrite.

Run a new provider-backed synthetic task:

sushi-browser-use demo --output "$HOME/browser-use-results/provider-$(date -u +%Y%m%dT%H%M%SZ)"

The task sends its synthetic page content and instructions to your provider. It is bounded to four agent steps with no provider retries. The model selects a native tool that reads the actual DOM and calculates the total, then uses the existing browser form action. Independent readback still requires the actual records and form POST. This demonstrates model-directed browser tools with native calculation.

Keep keys out of command arguments, task text, shell history and recordings. Histories, browser profiles and raw agent logs may contain page or task content; protect them as account data.

3. Inspect saved work

Each completed walkthrough includes extracted.json, submission.json, history.json and result.json. The result binds the actual saved artifacts with hashes. Failed runs remain available. To reread a completed task without contacting a provider or launching Chromium:

python3 -I -B /opt/browser-use/readback.py /absolute/path/to/the/completed/run

Use the actual directory printed by the walkthrough. For custom tasks, follow the request-file and explicit allowed-domain instructions in /opt/sushi/docs/README.md. Review the task, website and files you authorize. Domain navigation controls are not a network firewall or a secure isolation guarantee.

Persistence and recovery

Settings, task results and browser profiles remain on the root EBS volume across reboot. They are not automatically backed up; deleting the volume can remove them. Only the ubuntu account should operate the launcher. Chromium's sandbox remains enabled, and its debugging endpoint stays on the instance loopback interface.

If automatic preparation fails, inspect the service and journal through your authenticated administration session:

systemctl status sushi-browser-use-firstboot.service
sudo journalctl -u sushi-browser-use-firstboot.service

Inspect AppArmor and kernel logs if the browser cannot initialize its sandbox. Preserve failed output, correct the cause and use a new output directory. Do not disable the browser sandbox or make settings world-readable.

Original licenses and corresponding source

Browser Use 0.10.1 is packaged under its original MIT license. Application/dependency notices are under /usr/share/doc/browser-use; original wheels and source archives are under /opt/browser-use/wheels and /opt/browser-use/sources. Chromium's complete Debian copyright files, matching source archives and private compatibility-library sources accompany the package.

The original lxml license bundle remains in its installed dist-info directory. The lxml source and complete libiconv 1.18, libxml2 2.14.6, libxslt 1.1.43 and zlib 1.3.2 archives accompany it. The installed README documents its original source rebuild entrypoint. Original open-source terms govern those components, including their source-access, modification and relinking rights. Packaging/support charges do not replace those terms.

Version-specific upstream guide · Original Browser Use license