Transcription
Today, we're going to look at Manus, a general AI agent. I got early access, so I'll test it out and see what we can do. I wasn't sure what people wanted to see, but I made a few tests. I don't have unlimited access, so we'll keep it short.
Here's what I came up with: We'll try social media, logging into x.com. Can we research and write a tweet or post? Can we find other tweets and summarize them? I want to make a presentation on Anthropic MCP servers. Can it research, put everything together, and create a presentation, including images? Finally, I want to upload a CSV file, read it, and create graphs and an overview. That's our plan today using Manus.
This is new to me; I've only done a few tests. Let's look at the interface. It's pretty familiar: a collapsing sidebar with sessions, a text box, and a file upload option. There's a "standard" and "high effort" setting; I'll use standard. There are some examples. Let's start by asking, "Can you log into x.com?"
It seems to open a virtual machine or container. You can see "connected to data source," and "choose to take over the browser." Clicking "take control" opens a browser window. We go to x.com. Logging in was no problem; it's just an experimental account I use for AI testing.
After logging in, I asked it to research and write a post explaining "vibe coding." Manus is working, explaining logging limitations. It's searching, editing a file, and outlining its plan: visit websites, compile information, draft a tweet, and guide the user. It's not posting for us, but let's see.
It's visiting medium.com and other sites. So far, it's smooth. There's a "take control" option, a safety feature. It finished browsing and is compiling information and drafting the tweet.
It gives several options. I chose one and asked it to post it on x.com. It went back to the browser, found the post area, and posted the tweet. It wasn't super effective initially due to security limitations, but it eventually did it. The tweet was posted successfully.
Overall, it worked, but not perfectly. Let's move on to the presentation. I copied the prompt and pasted it into a new session: "Create a presentation on Anthropic MCP servers: what it is, how to use it, and why. Include images."
It went into the terminal, created a directory, and started researching. It visited Anthropic's page and documentation. It created research notes. Then it created an outline, collected images, and started creating slides. It used MDX format and added images. It reviewed, finalized, and saved the presentation.
The presentation is good, though not perfect. The image selection was slightly off, but the outline and content were good. It followed the steps well, except for a minor image search loop. For an early product, it was smooth.
Finally, let's upload a CSV file with API pricing data. I want an overview and graphs for input and output token prices of two models from each provider.
It's looking at URLs but missed one. It's browsing pages and gathering pricing information. It created a CSV with the pricing data. It ran into issues with OpenAI's Cloudflare protection, but found a workaround. It created a Python script, installed seaborn, and combined the price data.
It generated an MDX file with HTML for the graphs. The graphs look good, showing a comparison of API pricing and token pricing. GPT-4 Turbo is expensive, but Claude 37 is also pricey.
We covered everything. Manus looks good, but this is an early test. It's not a sponsored video; I just got early access. I'll keep testing it. If people like this video, we might do another one. It's cool to see AI agents evolving incrementally. I think Manus is from a Chinese company, and they’re doing well. Hope you enjoyed it!