Lodestar

Inspiration

With the growing economy of AI agents interacting with the web, we noticed a glaring inefficiency with projects like Claude Computer Use

Imagine an AI agent trying to place an order on an e-commerce website, it needs to gain access to the website, parse the UI layer, understand how to navigate the website, and then hopefully finish the action.

The question that we are asking building this company is why? Why are these projects increasing the cost for the hosted website’s infrastructure as well as the cost for the developer building these tools

Because every parse costs input tokens for a result that is at max, satisfactory but usualy subpar and this is repeated for every user who asks more or less the same query.

The reason is pretty simple, its because, right now, in 2025, there are no tools that give out an universal action layer for websites and That is what Lodestar is building

What it does

Lodestar parses a website once and creates reusable tools for performing actions. One of the key benefits of us doing this is that it makes the action layer accessible to both small browser based models but also large language models and it provides a medium to get consistent and accurate results

Sponsor Challenges

Fingerprintjs was used in a way to track the users who generate the queries to preload the websites they might ask for but also for potentially anonymising their data and give it to interested marketing companies, basically tracking the bot and aligning it to a user when they change bots.

Twilio for sending messages when the agent is about to perform like a financial transaction or anything that would require a human review