today was quite a fun filled one. and a little stressful.

now a while back i came across the claude code embedded browser feature, and i was beyond intrigued at how it worked. watching it navigate pages, interact with sites, just do things visually, it stuck with me.

and as a huge fan of amazon q, the very underrated agent that does a lot of work, as good as claude, i kept thinking, why doesn't q have this.

the blind agent

over the course of about a year, i've been using amazon q for a lot of local projects. it's genuinely good. but the one downside that always got to me was that it was literally blind.

unlike claude, which could access the internet seamlessly, amazon q wasn't allowed to. i don't know, maybe it was a security decision the company took because of how lowkey but powerful the model actually is. but whatever the reason, it couldn't see the web.

and that was genuinely painful to work around. having to explain what was happening in the browser, copy-pasting errors, describing UI states, it was tedious. claude had a visual understanding of what was happening on screen. amazon q didn't.

that's why i called it blind at the beginning.

so i built it eyes

i built a local mcp, model context protocol, that lets amazon q browse the internet right from the terminal.

the same embedded browser tech that claude uses. i basically gave amazon q the same capability.

the process was actually pretty simple. i asked claude to give me an initial overview of how its embedded browser worked. it gave me around 80 tools. literally built the entire v1 of the mcp.

then i took that v1 to amazon q and asked it to help amplify the system as a whole. and here's the thing, amazon q knew what it was lacking in some way. it enhanced the tools, added more features, tightened things up.

the difference in how they handle credentials

one thing i noticed, claude has this feature where you can give it a URL, provide credentials, and it'll just go there and navigate. no questions asked.

amazon q is different. it's all about safety. it'll say things like "i can't help with that" or "don't paste credentials in the conversation", which is a normal safety precaution, i get it.

but it won't just leave you stuck. it'll advise you to create an env file with a structure like:

and then it just reads the env, logs in, and asks what you want it to do next. honestly, that's the right way to do it.

on claude

i'll be honest, i'm not a claude fan. the 5-hour reset thing is annoying. the excessive guardrails are worse. messages getting flagged, especially anything related to security audits or anything it's uncomfortable doing. the classifier is overzealous and it makes certain kinds of work genuinely hard to do.

i never even paid for it. a friend shared his subscription with me and it was hellish. some messages just get flagged for no good reason. and now they're adding more stuff like the watermark situation.

this isn't me saying claude code is bad or that this is a promo for amazon q. i'm just saying the system has been frustrating lately and i'd rather use a tool that doesn't fight me.

where it's at now

this is v2. i'm still building. let's see how it goes from here.

but the short version is, i built a means for my previously internet-blind ai agent to actually access the web and do things. it can navigate, it can read pages, it can interact with sites, all from the terminal, all through amazon q.

it's not perfect. but it works. and that's enough for today.