MediaUse - Turn Web Tasks Into Reusable Workflows

The core idea is simple: if a browser task is already known and repeatable, the agent should not re-solve it every time. Today, most AI browser agents operate like this on every execution: open page → inspect DOM → locate elements → reason about UI → execute click/type → handle edge cases → repeat This works for unknown or exploratory tasks, but it becomes inefficient and expensive for workflows that are already understood. MediaUse introduces a different abstraction: **semantic CLI workflows for the web**. Instead of repeatedly reasoning about page structure, the agent calls a stable command interface such as: search.hot get.detail post.feed account.health The agent decides *what it wants to achieve*, while MediaUse executes the actual browser automation locally through a CDP/browser runtime and site-specific plugins. This shifts execution from “token-per-step reasoning” to “function-level invocation”. It is particularly effective for fixed, high-frequency workflows such as: * searching known websites * reading structured lists and feeds * opening detail pages * submitting forms * posting content * checking account status * replaying previously validated workflows In these cases, the model is no longer required to inspect the DOM or make per-step decisions. It only selects a semantic command and receives structured outputs. This reduces token usage during execution close to zero for most workflows, while improving determinism and latency. A key advantage of this approach is separation of concerns: * AI handles intent, planning, and selection of workflows * MediaUse handles execution, reliability, and site-specific logic All low-level details—selectors, fallbacks, login flows, anti-popup handling, retries, and output schemas—are encapsulated inside versioned site plugins. These plugins are testable, reusable, and deterministic. We currently support a growing set of production-ready site plugins, including: arxiv, bilibili, bloomberg, chatgpt, douyinv2, fifa2026, google, hackernews, huggingface, instagram, reddit, tiktok, twitter, weibov2 Beyond execution, we have also built an **agent-based workflow system for generating and maintaining site plugins automatically**. This is a critical part of the system: websites change frequently, which is one of the biggest failure points in traditional browser automation. MediaUse addresses this by using an agent workflow that can: * generate new site plugins from a URL * validate workflows through execution * update broken selectors and flows automatically * maintain versioned compatibility over time In practice, a developer can turn any website into a CLI-based agent interface in under 30 minutes. Once a workflow is created, it becomes a reusable interface that can be executed at very low cost and high reliability. Importantly, this is not limited to read-only automation. The local runtime also supports authenticated and interactive actions, meaning agents can perform real user operations such as posting, messaging, submitting forms, and interacting with web applications. Overall, MediaUse is not trying to replace general-purpose browser agents. Those remain essential for exploratory tasks and unknown environments. Instead, we focus on a complementary layer: > turning known web workflows into deterministic, reusable, low-cost execution primitives. I believe this is the missing infrastructure layer for production-grade AI automation on the web.