browser automation

Bilingual

Agent-Browser – Browser Automation for AI
2 months ago
- agent-browser is a browser automation CLI designed for AI agents, featuring compact text output to minimize token usage.
- It uses a ref-based accessibility tree for deterministic element selection, requiring only 200-400 tokens per snapshot.
- Supports 50+ commands for navigation, forms, screenshots, network, storage, tabs, frames, and debugging.
- Includes built-in observability tools: video recording, streaming, debugging, profiler, and diffing.
- First-class support for modern apps with network control, React, Web Vitals, Next.js, Vercel, and event extension workflows.
- Stateful capabilities: sessions, profiles, auth state, cookies, storage, proxy, and security controls for long-running agents.
- Cross-platform: macOS, Linux, Windows with native binaries; works with AI coding agents like Claude Code, Cursor, GitHub Copilot.
- Client-daemon architecture: Rust CLI parses commands and communicates via Chrome DevTools Protocol; daemon persists and auto-exits after idle timeout.
How to build agent infrastructure yourself
2 months ago
- Every web-touching agent requires a browser.
- At scale, infrastructure includes five layers: warm pools, sandboxing, identity sites trust, session replay, and a model gateway.
- These layers are all buildable, but each represents a full-time job.
Show HN: Browser Tools SDK – an optimal browser harness for agents
2 months ago
- Six tools enable any AI agent to open a real browser, read the page, and act using Playwright.
- The agent reads a compact page snapshot and runs needed Playwright code, reducing overhead and cost.
- Returns a compact accessibility tree with stable refs, avoiding raw HTML.
- Each Playwright execution returns a compact snapshot diff showing changes on the live page.
- Includes adapters for AI SDK and Pi, with framework-neutral base tools for custom agent loops.
- Supports multiple providers: LocalBrowserProvider, Libretto Cloud, Browserbase, Kernel, and Steel.
- Two core tools are browser_snapshot (to read the page) and browser_exec (to run Playwright); four others manage sessions.
- The agent uses a small tool set and executes Playwright code, giving access to the full Playwright API.
- The package is MIT licensed and available in the Libretto repository.
- Install the package without changing your current model or agent framework.
Self-testing AI harness finds its own bugs
3 months ago
- Every morning at 9 AM, Omnideck stress-tests its browser tools on 50 real websites across categories like e-commerce and news.
- When tools fail, Omnideck uses diagnostic tools to identify root causes, such as DOM issues or race conditions, and validates them against its own source code.
- Over a week, this routine found 14 bugs across 35 sites, with examples including content selection logic and modal handling errors.
- Bugs are filed as GitHub issues with proposed fixes, creating a self-improvement loop that enhances tool reliability for future tasks.
- The process leverages existing capabilities like browser tools and code analysis to systematically reduce web chaos and improve performance.
1Password for Claude: Give Claude access without giving up your credentials
3 months ago
- AI agents are evolving to perform actions on behalf of users, necessitating a new security model for credential access.
- 1Password for Claude enables secure credential usage without exposing passwords or one-time codes to the AI model, using a zero-exposure architecture.
- Users approve credential requests via biometric authentication, and 1Password injects credentials directly into web pages, ensuring secrets remain encrypted and controlled.
- Agentic Mode in the 1Password browser extension locks down the interface when AI agents take control, restricting access to only approved credentials for the current task.
- The solution supports various use cases, such as personal tasks (e.g., redeeming Audible credits) and business operations (e.g., accessing Stripe dashboards).
- 1Password for Claude is part of a broader ecosystem for securing AI agent access across browsers, IDEs, and workflows, with availability for Mac users on different plans.

browser automation

Agent-Browser – Browser Automation for AI

How to build agent infrastructure yourself

Show HN: Browser Tools SDK – an optimal browser harness for agents

Self-testing AI harness finds its own bugs

1Password for Claude: Give Claude access without giving up your credentials