The model hadn't tweeted yet, but it already knew: Sai turns the browser into a release radar
Bhavy built an AI model tracker with Sai MCP: it periodically opens a real browser to check changelogs and sends a notification when there's an update. Being first is only the surface; the community is more concerned with how it distinguishes new models from edited logs.
Bhavy posted an AI model tracker he built himself on X. The core usage is not complicated: use Sai MCP to schedule the task, and Sai opens a real browser through computer use, checks the changelog, and sends him the details once it finds a new model. The task runs on a schedule, so he can know before the lab tweets. The original tweet is here: Bhavy's tweet, and the Sai entry point is: sai.simular.ai.
A changelog is a product update log. Model teams sometimes update their docs or logs before posting on social media. For people watching model releases, manually refreshing pages is dumb but effective. Sai's approach is to have the model open a browser, click through pages and read changes like a human, then repeat on a schedule.
Demo video:
It looks like a first-scoop tool, but people in the replies are already pouring cold water on it. Some users asked whether it would trigger bot detection on changelog sites, while others asked how it distinguishes a new model release from an edit to an existing entry. The latter question is more critical: simply telling you something changed isn't enough; it's better to include a date diff and the official URL so people can verify before sharing. Others asked about token costs.
This also explains the real value of similar tools. Being faster than the official account isn't the hardest part; the hard part is making changes verifiable, traceable, and low in false positives. Once model release information enters the distribution chain, being half a beat fast can easily become half a beat wrong. What's interesting about the Sai case is that it pulls computer use out of demo videos and into boring scheduled tasks: open a browser, read logs, send a message. The workflow isn't cool, but it's close to real work.
For ordinary users, it can be understood as a reminder that refreshes webpages by itself; for readers with a development background, the focus is on MCP integration, computer use execution, and scheduled orchestration. As for whether it can really run reliably on changelogs, that depends on bot detection, page structure changes, and diff strategy. Bhavy's tweet gave the result; the community's follow-up questions gave the cost.
发布时间: 2026-11-03 18:28