Opensteer
A compact Python interface for agents to operate local and cloud browsers.
- Source repo
- steerlabs/opensteer
- Stars
- ★ 260
- Last updated
- today
- License
- MIT
- Primary language
- TypeScript
- FA score
- 54/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platform
- Cost
- Free, no paid service needed
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- Agent developers who need Chrome or Edge navigation, clicking, typing, and page inspection from Python commands.
- Not a fit if
- Teams that cannot allow agents to execute local Python or JavaScript
- Users expecting Opensteer Cloud access without an API key
- Workflows that require read operations to provision cloud browsers automatically
- Website
- opensteer.com ↗ set on GitHub
- Maintenance
- Commits in the last 30 dayslast commit 2026-11-20latest release desktop-v0.1.685 (2026-11-18)+6.7 ★/day over 7 days
- Source review
- 54/100 · Major gaps
What does this agent do, and when should you use it?
Opensteer is a browser automation framework that gives AI agents a small Python command surface and a set of importable helpers. It handles CDP transport, local and cloud attachment, named sessions, tabs, screenshots, JavaScript execution, clicks, typing, and scrolling. Users can run short snippets through opensteer -c or import functions from opensteer.helpers into reusable Python code. A project directory acts as a specialized harness, supplying AGENTS.md, selectors, data, API clients, and custom functions for a particular workflow. Opensteer can attach to local Chrome or Edge and, when OPENSTEER_API_KEY is configured, to Opensteer Cloud browsers with explicit session and transport controls.
An automation run begins with a Python snippet passed to opensteer -c. The code can create, select, or enumerate browsers through open_browser(), use_browser(), and list_browsers(), then open pages with new_tab() and wait using wait_for_load(). It can inspect a page with page_info(), navigate with goto_url(), click with click_at_xy(), enter content through type_text(), scroll, manage tabs, capture screenshots, execute JavaScript through js(), or invoke raw CDP methods. When several browsers are active, the caller retains each returned browser handle and distinguishes sessions with readable labels. Workflow-specific modules such as tools.py and selectors.py are imported from the working directory, allowing a command to produce page information, screenshots, or custom function results. Local App execution can request auto, inapp, local, or cloud, while Cloud execution accepts only auto and cloud.
- Agent developers who need Chrome or Edge navigation, clicking, typing, and page inspection from Python commands.
- Support teams building a dedicated harness that wraps customer-page actions in functions inside
tools.py. - QA teams that need JavaScript evaluation, screenshots, and raw CDP access while investigating web behavior.
- Developers running concurrent research, work, or site-specific browser contexts through named sessions and explicit handles.
- Sales, data-entry, or research teams that want separate selectors and workflow data while sharing one browser-control layer.
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Install / run
curl -fsSL https://opensteer.com/install.sh | sh
- Environment variables
OPENSTEER_API_KEY
How do you install or deploy this agent?
Run the documented installer from a shell:
curl -fsSL https://opensteer.com/install.sh | shVerify the installation and attempt the first page inspection:
opensteer --doctor
opensteer -c "print(page_info())"If local attachment is not ready, start the setup flow. It guides Chrome or Edge through enabling remote debugging for the current profile:
opensteer --setupOPENSTEER_API_KEY is required only for Opensteer Cloud browser access. The source does not provide a specific shell command for assigning the key.
How do you use this agent?
Open a page, wait for it to load, and print its page information:
opensteer -c "new_tab('https://example.com'); wait_for_load(); print(page_info())"Import browser helpers into reusable Python code:
from opensteer.helpers import goto_url, js, click_at_xy, type_text, wait_for_loadKeep the returned handle when operating more than one browser:
opensteer -c "linkedin = open_browser(label='linkedin', profile='fresh'); linkedin.new_tab('https://linkedin.com')"After configuring OPENSTEER_API_KEY, create or attach to a cloud browser:
opensteer -c "open_browser(label='research', profile='fresh'); new_tab('https://example.com')"opensteer -c "browser = open_browser(label='work', profileId='bp_...'); print(browser.page_info())"List and select existing browsers without creating another one:
opensteer -c "print(list_browsers())"
opensteer -c "research = use_browser('research'); research.new_tab('https://example.com')"A locally executing agent can explicitly request a cloud browser:
opensteer -c "open_browser(label='research', profile='fresh', location='cloud')"A specialized harness can define a local workflow function and import it directly from the CLI:
# tools.py
from opensteer.helpers import goto_url, js, type_text, wait_for_load
def open_customer(customer_id):
goto_url(f"https://support.example.com/customers/{customer_id}")
wait_for_load()
return js("document.title")cd support-harness
opensteer -c "from tools import open_customer; print(open_customer('cus_123'))"What are this agent's strengths and limitations?
- One helper surface covers CDP, local and cloud attachment, sessions, screenshots, JavaScript, and common page interactions, reducing the need for a custom browser bridge.
- The
opensteer -cinterface suits short operations, whileopensteer.helperssupports reusable Python modules as workflows mature. - Browser infrastructure is separated from workflow assets: Opensteer owns transport and generic interaction, while each project retains its selectors, data, instructions, and business functions.
- Named browsers and explicit handles support concurrent browser contexts, with
modecontrolling whether a logical session is reused or intentionally created.
- Local attachment depends on Chrome or Edge remote-debugging setup and may require an additional
opensteer --setupstep. - Cloud browsers require
OPENSTEER_API_KEY, and Cloud execution accepts onlyautoandcloud, rejecting device-only locations. - Domain-specific automation is not supplied out of the box; adopters must maintain their own instructions, selectors, data files, and Python functions.
- The interface can execute Python, in-page JavaScript, and raw CDP calls, so adopters must define an appropriate security boundary for agent execution.
- The source documents no native adapter for ChatGPT, Codex, Claude, or their APIs.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Opensteer This agent | 54 · Major gaps | CLIFree | ★ 260 | today | TypeScript | — |
| Obscura Headless Browser | 68 · Some gaps | CLIFree | ★ 37k | 1d ago | Rust | Claude.ai |
| Open Browser | 39 · Major gaps | CLIFree + model costs | ★ 9.6k | 7mo ago | TypeScript | ChatGPT · Claude.ai · OpenAI API · Claude API |
| Browser4 Agentic Browser | 65 · Some gaps | CLIFree + model costs | ★ 1.2k | 5d ago | Kotlin | OpenAI API |
How does FollowAgents rate this agent?
Why each dimension lost points
The materials state that files and credentials remain local, describe password and token boundaries, and require explicit OAuth authority grants. Browser control, desktop control, and capabilities such as agent shell access remain powerful, while the packet does not show the complete permission model, so least privilege and external effects score 1. Setup and integrations include user configuration or consent, earning 2. Rollback evidence covers only some recovery procedures, and attribution consists mainly of a private security-reporting route and MIT copyright notice, so each scores 1.
CI, test commands, and safe failure classifications are documented, but no execution results are included, limiting what can be verified. The README says there will be no cloud-hosted product, while package.json describes durable cloud agents; this product-position mismatch earns 1 for self-consistency. Versions and toolchains are declared, but availability guarantees for external agents, browsers, and services are thin, earning 1 for dependency availability. The OAuth document describes actionable failure handling, earning 2.
The product clearly targets people who want to steer agents running on their computers from a phone, with web, desktop, mobile, weak-network, and multiple-agent scenarios. The documents also state authority boundaries and early-alpha limitations, and skill descriptions provide fairly clear triggers. Platform and dependency requirements are concrete, but the full compatibility range across environments is limited, so these criteria score 2.
The repository organizes material through the README, topic-specific documents, and skills, and provides concrete setup steps. However, the current process is a source build rather than a simple end-user installation guide. Naming and package descriptions differ on cloud positioning; examples and FAQs are limited. Alpha status and security limitations are disclosed. The complete MIT license text earns 3. A version is present but the packet contains no change history; maintenance responsibility is evidenced mainly by the private vulnerability-reporting route, so those criteria score 1.
The product goal and outputs are understandable: run agents on a computer and steer them from other devices, across coding agents and browser or app work. Local operation and multiple clients offer practical value, but the cost-benefit case rests mainly on claims such as avoiding API bills. Alpha status, source-build requirements, and platform toolchains reduce current usability, so cost-benefit scores 1.
Some README claims about setup, security, and product behavior have matching support in AGENTS.md, skills, and the OAuth document; CI configuration and source entry points provide traceable leads. The materials distinguish current behavior, limitations, and future plans, but include no execution evidence and contain some contradictory positioning, so each criterion scores 2 rather than full marks.
- The project identifies itself as early alpha, and the README says users currently build from source; treat the documents as neither a security guarantee nor proof of successful execution.
- The README says a cloud-hosted product is not planned, while package.json describes a cloud-agent workspace; product-positioning evidence is inconsistent.
- Full-access agents, signed-in browser sessions, saved logins, and remote device connections involve sensitive data or broad authority; review the concrete grants and data flows before use.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
desktop-v0.1.685 Opensteer 0.1.685 (Alpha)
FAQ
Is Opensteer Cloud mandatory?
OPENSTEER_API_KEY is configured.Will a read helper automatically create a browser?
open_browser() first or select an existing browser with use_browser().What is the difference between mode='reuse' and mode='new'?
reuse is the safe default for reusing a logical session. new deliberately creates another logical session. Labels are human-readable, while cloud session IDs and run scope are platform-owned.How should separate sites or business workflows be organized?
Is hosted browser pricing documented?
Is Opensteer open source?
Is Opensteer safe to use?
Is Opensteer still maintained?
Who makes Opensteer?
What is Opensteer's official website?
Thanks — we use this to improve the page.