Show HN: Argus,为那些编码代理(AI Agent)交付速度快于 QA 测试团队的开发团队提供的代理化 QA 工具。
Show HN: Argus, agentic QA for teams whose coding agents move faster than QA

原始链接: https://github.com/argus-testing/argus

Argus 是一款智能视觉 UI 测试工具,无需手动编写脚本,也无需维护易失效的选择器。只需描述测试意图,Argus 就会利用自主多智能体系统像真实用户一样浏览、点击并与您的 Web 应用进行交互。 Argus 基于 Playwright 构建,通过五个专业智能体——验证者(Validator)、理解者(Comprehender)、探索者(Explorer)、策略师(Strategist)和执行者(Executor)协作,完成 UI 映射、测试计划制定及实时执行。该系统提供透明的逐步推理过程,并附带时间线、截图和结构化的通过/失败报告。 主要优势包括: * **零脚本编写:** 无需维护固定脚本,智能体会自动适应您的 UI。 * **快速设置:** 可立即用于本地主机(localhost)、私有网络及任何 HTTP(S) 目标。 * **注重隐私:** 采用源码可见、本地优先的方案,使用 SQLite 存储;您的数据和截图绝不会离开您的机器。 * **完全可视化:** 通过基于浏览器的仪表板实时查看智能体的工作过程。 Argus 对个人及小团队(100 人以下)免费,专为无缝本地优先部署而设计。如需了解更多信息,请访问 argustest.com。

抱歉。
相关文章

原文

AI agents that test your UI like a real user — no scripts to write, no selectors to maintain.

Website

Argus is a visual UI testing agent. Describe a test, point it at an HTTP(S) page, and watch it inspect the page in an isolated Playwright browser context. Runs, timelines, screenshot references, and structured reports are captured automatically.

Argus finds the bugs you didn't write tests for. Point it at a page, describe what "working" looks like, and an autonomous agent explores your UI the way a real user would — clicking, typing, scrolling — then hands you a structured report with screenshots and a timeline. No test scripts to maintain, no flaky selectors to babysit.

Argus dashboard

  • Agentic, not scripted — the agent reasons about the page and adapts, it doesn't replay a fixed script
  • Fits your stack — built on Playwright, works against localhost and private-network apps
  • Zero setup ceremony — start testing in minutes

The fastest way to run Argus: no install, no API keys to manage, nothing to self-host.

Get started at argustest.com.

A run isn't a single model call — it's five agents handing off to each other, and you can watch each one work in the live run view:

  1. Validator — checks the target URL and test description are actually testable before a run starts.
  2. Comprehender — reads the test description and breaks it into distinct test cases.
  3. Explorer — crawls the app first, mapping out pages and the actions available on each.
  4. Strategist — turns the map and test cases into a concrete step-by-step plan.
  5. Executor — runs the plan in a real browser: navigating, typing, clicking, and confirming outcomes as it goes.

Nothing happens in a black box — each agent streams its reasoning as it works, down to individual actions like "Navigating to /companies" or "Confirming 'Airbnb' is on the page," so you see the app get mapped, the plan get built, and the test get executed, live.

Describing a test

A run in progress

A completed run report

Integrations

A test is just a description of intent — Argus figures out how to interact with the page. For example, pointed at Y Combinator's site:

Target: https://www.ycombinator.com

Test the startup directory.

1. From the homepage, navigate to the companies/startup directory.
2. Search for a well-known YC company by name (e.g. "Airbnb") and confirm it appears in the results.
3. Filter the directory by a specific batch (e.g. "Winter 2024") and confirm the listed companies update to match.
4. Open a company's profile from the results and confirm its name, one-line description, batch, and website link all render correctly.
5. Navigate back to the directory and confirm the search/filter state behaves as expected — either preserved or reset, whichever the page is designed to do.
6. Resize the viewport to 375px width and confirm the nav collapses into a mobile menu, and the directory list stays scrollable and usable with no overlapping elements.

Fail the test if the known company doesn't appear in search results, if the batch filter doesn't actually filter the list, if a company profile is missing expected fields, or if the mobile layout breaks.

Argus runs this like a person would — clicking through the flow, reading the page to judge success or failure — and returns a timeline, screenshots at each step, and a pass/fail report with the reasoning behind it.

Argus is source-available and fully local-first — SQLite storage, no telemetry, your data and screenshots never leave your machine. Prefer to self-host? Follow the steps below.

cp .env.example .env
# Set GEMINI_API_KEY in .env
uv sync --dev
uv run playwright install chromium
cd frontend && npm install && npm run build && cd ..
uv run uvicorn argus.app:app --reload --env-file .env

Open http://localhost:8000. For frontend hot reload, run npm run dev in frontend/ alongside Uvicorn and open http://localhost:5173.

Configuration is environment-only:

Variable Default Purpose
GEMINI_API_KEY — Required for real execution
GEMINI_MODEL gemini-2.5-flash Gemini REST model
ARGUS_DATA_DIR data SQLite and screenshot directory
ARGUS_HEADLESS true Playwright browser mode
ARGUS_RUN_TIMEOUT 300 Run timeout in seconds

Argus accepts normal HTTP(S) targets, including trusted localhost and private-network apps. It rejects credentials and sensitive query parameters in target URLs, and never stores provider secrets, typed browser values, or inspected page content. The settings screen only shows whether provider configuration is present.

cp .env.example .env
# Set GEMINI_API_KEY in .env
docker compose up --build

The UI is available at http://localhost:8000 and persistent data is written to ./data.

uv run pytest
uv run ruff check argus tests
uv run pyright argus tests
cd frontend && npm run typecheck && npm run lint && npm run build
  • argus/runtime: provider-neutral agent, message, tool, and session boundary
  • argus/providers/gemini.py: raw httpx Gemini REST/SSE adapter (no provider SDK)
  • argus/pipeline.py: planning, one browser agent, evidence capture, and reporting
  • argus/store.py: SQLite runs, events, screenshots, and reports
  • argus/app.py: REST API, reconnectable per-run WebSockets, and built UI serving
  • frontend: dashboard/composer, live session, history, report, and read-only settings

API docs are available at /docs. All data is local; there is no authentication or multi-user isolation in this release.

Argus is source-available under the Argus Source-Available License 1.0 (ASAL-1.0).

  • Individuals & Small Teams (< 100 members): Free to use, modify, and self-host for both commercial and non-commercial purposes.
  • Enterprises (100+ members): Requires a commercial license. Contact us at [email protected] to get set up.
  • Releases automatically convert to the MIT License after 3 years.
联系我们 contact @ memedata.com