ECOCREATE TECHNOLOGY PTE. LTD., a Singapore-based company, open-sourced two agent skills on GitHub on May 14 that address one of the persistent gaps in AI agent infrastructure: interacting with the live web without getting blocked, burning excessive tokens, or breaking when sites change.
The two skills, browser-act and browser-act-skill-forge, work with Claude Code, Codex, OpenClaw, Hermes, and Cursor. According to the announcement, agents using browser-act report 90% fewer error-and-retry loops and 93% reduction in token consumption compared to hand-written scrapers, because output returns as structured JSON instead of raw HTML.
The Problem BrowserAct Targets
AI agents can reason well but act poorly on the web. Bot-detection services like Cloudflare, DataDome, and hCaptcha guard more than 40% of the world’s top 10,000 sites, according to the GlobeNewswire release. Standard browser automation tools (Puppeteer, Playwright) get detected on the first attempt. A single page’s raw HTML can run thousands of tokens, inflating costs. And every new site requires custom scraping code that breaks the moment the site redesigns.
Earlier solutions addressed pieces of this: converting web content into AI-friendly text, building search APIs for agents, adding AI features to consumer browsers. None shipped a complete, open execution layer that agents can call and extend themselves.
browser-act: A Stealth Browser for Agents
The first skill is a runtime that exposes simple verbs (navigate, click, type, extract, login, screenshot) as tool functions. Three layers of isolation distinguish it from generic browser automation.
Fingerprint isolation. Every session gets randomized canvas, WebGL, audio, and navigator properties. Even in headless mode, bot-detection services cannot distinguish it from a human session. Network isolation. Each instance runs behind its own residential IP with auto-rotation for long pipelines. Session isolation. Each run is a fully independent identity with its own fingerprint, IP, and cookie jar. Two runs of the same skill cannot be correlated, per the GitHub documentation.
Fingerprint isolation, stealth browsing, and CAPTCHA solving (hCaptcha, reCAPTCHA, Cloudflare Turnstile) are free and open-source. Residential proxy routing across 195+ countries is an optional paid add-on.
A feature called Chrome Takeover lets agents connect to an existing Chrome session, inheriting authenticated state across Gmail, LinkedIn, Amazon, or internal dashboards without credential handoff.
For regulated workflows, a remote-assist feature allows human intervention mid-task. If an agent hits an SMS verification or high-value confirmation step, it generates a link and pauses. A human reviewer completes the step in the same browser session (same fingerprint, same IP, same cookies), then hands control back. The target website sees no session switch.
browser-act-skill-forge: Agents That Build Their Own Tools
The second skill addresses a different bottleneck. When an agent encounters a website nobody has automated before, someone has to write fresh scraping code, or the agent regenerates brittle code on every run.
Skill-forge automates that process. Point it at any website, and it produces a reusable skill the agent (and anyone else) can call going forward. On first visit, it opens a real browser, explores the site, finds API endpoints where possible, and falls back to DOM patterns. The resulting skill is reusable across agents and platforms.
The GlobeNewswire release describes it as solving the “explore once, reuse forever” problem, contrasting it with the current pattern where “whatever an agent learns about one website can’t be saved, can’t be shared, and breaks the moment the site redesigns.”
Why It Matters for Agent Infrastructure
BrowserAct is notable less for what it does (browser automation is not new) than for how it packages it: as platform-agnostic agent skills that slot into existing frameworks without custom integration work. Installation is a single command (npx skills add browser-act/skills --skill browser-act), per the GitHub repository.
The release also reflects a broader infrastructure maturation pattern in the agent ecosystem. Agents are moving from custom-coded capabilities toward shared, maintained components. Red Hat launched an agentic skills repository at Red Hat Summit 2026 last week. OpenClaw’s own skill marketplace continues to grow. BrowserAct adds a web-access layer to that emerging shared infrastructure.
The production use cases cited in the announcement range from job application automation (100 applications in 10 minutes) to competitor price monitoring across geographic regions to multi-account management with cross-contamination prevention. Whether those claims hold at scale will depend on how the open-source community pressure-tests them.