github.devotd.
NewAI agents

hyperbrowserai/hyperagent v1.0.0 — Natural-language Playwright browser automation

HyperAgent is a browser agent that carries out Playwright-based browser tasks from natural-language commands and also provides structured data extraction and replay of recorded tasks.

Repositoryhyperbrowserai/hyperagent ↗

HyperAgent is a browser agent that carries out Playwright-based browser tasks from natural-language commands and also provides structured data extraction and replay of recorded tasks.

GitHub description: AI Browser Automation

stars
★ 1.6K
Forks
225
Language
TypeScript
License
NOASSERTION
Latest release
v1.0.0
Read original source ↗View on GitHub ↗

hyperbrowserai/hyperagent is a browser agent that adds natural-language commands to Playwright. Judging by the GitHub releases, the latest version is v1.0.0, the repository description is "AI Browser Automation", and the primary language is TypeScript. The star count is 1598, and the repository was created on April 14, 2025.

The README describes the tool as AI-enhanced, Playwright-based browser automation. A single task is handled by page.perform(), a multi-step sequence of tasks by page.ai(), and page.extract() pulls out structured data. It states that plain Playwright can be used again when AI is not needed.

On top of this come cloud scaling through Hyperbrowser, an MCP Client, and Stealth Mode. A task recorded once can be replayed exactly as is, without any LLM call. The basic form of the tool is that the target of browser automation is specified in natural-language sentences instead of code, and then executed.

That said, some points remain unverified. The license shown by GitHub is NOASSERTION, so the reuse conditions have to be checked separately, and the release notes for v1.0.0 show no confirmable changes, meaning what went into this version has to be looked at directly on the repository's releases page. The evaluation record only has a momentum score of 0.0, with the remaining entries blank. On the evidence of the features listed in the README alone, it is close to being within reach of trying out the use case of bundling multi-step web tasks in natural language and running them, and any broader judgment would be premature.