Introducing the /browser-trace skill, Give your agent 100% observability into its browser: dump network requests, DOM content, screenshots, and CDP logs into a searchable filesystem. Great for reverse engineering, autoresearch loops, and monitoring the situation ™️.

Get started now!!

AI can doomscroll now all my friends are cooked

we're toast

me at work

@kylejeong This is very cool

@kylejeong let me know if you give it a try! this is really good for creating data for auto research loops

Get the best observability on our dashboard:

why have i watched this video for 2 hours

this is the missing piece. Had a playwright automation fail silently on 3 different staging envs last month. Network request visibility plus dom snapshot would've caught it in minutes instead of hours. The CDP logs thing especially, that's where the real signals live.

CDP = the truth

Isn't this also available in chrome-devtools-mcp or am I mistaken?

devtools MCP can drive your browser, but this skill can attach to any running browser while another process is controlling it, and help you determine exactly what happened at the browser, network, and content layers

What is the difference with devtools MCP?

devtools MCP can drive your browser, but this skill can attach to any running browser while another process is controlling it, and help you determine exactly what happened at the browser, network, and content layers

Does it bring a new layer on top of it? Like driving internal APIs or something else? I'm not sure what is the value there.

It brings o11y. So you can drive the browser with devtools MCP, set up browser trace as a audit trail for what happened on the browser when you drove it with the devtools mcp. It creates a parallel, read only websocket connection to the browser

Thanks!

what's the difference of just using chrome devtools mcp/cli?

You can use this with any browser automation tool to actually understand what happened while the tool was driving the browser. At every layer. The skill then organizes the data for you and informs your agent how to traverse it

Sounds useful! Will give it a go

hell yeah!

A bunch of these coming out lately. Will help if you could also highlight how it is different and if better than others 1. Codex in app browser 2. Browser use guys 3. Gstack browser -> I think cdp based too 4. Claude in chrome (probably least favored?) 5. Google chrome dev tools 6. Definitely missing a dozen more

You can attach browser trace to any running brower as it your automation, agent, or even a human drives it. Then from the o11y data, you can build more robust automations, power auto research loops, or just understand exactly what transpired on the browser in general. This is meant to be used in conjunction with other tools!

So sick

let's go!!!

browserbase 🤝 tiktok

Agents post videos for agents to doom scroll 💀

it’s agents all the way down

ayo what was that last vid

@kljukusa chill bro it was @shreypandya's algo

@shreypandya Sounds like someone was using the context api a bit too much…

This is a game-changer for debugging agents. Full traces, screenshots, and that searchable HTML report? Exactly what we needed. Shipped perfectly

cdp logs as a filesystem is the correct abstraction.

woo CDP logs as a searchable filesystem is the move, that’s where silent failures hide.

CDP = the truth

this is great, will try it out

this is sickkk

🤝🤝

브라우저 추적 이슈 디버깅할 때 진짜 유용할 듯

can confirm

这玩意儿牛啊，以后让AI自己扒网页写报告，咱躺平喝咖啡就完了。

does this work for arc browser?

it works with any chromium based browser!

cool, thanks!

This is exactly what we needed for debugging agent browser interactions. We've been using CDP logs manually but having it auto-dumped into a searchable filesystem changes the game for reverse engineering workflows.

if your agent can do it then no reason to do it yourself!

Nice, helpful to have the tools to do this investigation local via coding agent as opposed going on a platform and manually looking for shit

yupp. who even reads logs without an agent these days?

This is awesome, we've built something inhouse on similar lines for logs, network trace and screenshot, as the cleanup step, whenever our automations fail. But I wasn't aware we could attach to an existing session from outside with read only CDP.

CDP permits multiple connections! Having a separate parallel connection is the o11y cheat code

browser-trace gives agents full observability: network requests, DOM, screenshots, CDP logs into a searchable filesystem. This is the missing piece for debugging autonomous browser agents. Without it, agent failures are black boxes.

100% browser observability for your agent. The black box problem of agentic browsing just got solved.

bingo

right? 😅

How did it do on discerning the content it scraped contextually?

that's up to your agent !

response bodies make this multi-GB/hour on a busy SPA. searchability only helps if you filter at trace time. grepping 4GB of CDP logs is its own problem.

@yourkaisensei its not just a grep. the skill ships with .mjs files, a cdp summary, and query recipes to make searching easy

fair, recipes kill the easy 80%. the bite is incident triage when you don't yet know what to look for. that's still raw grep volume.

@yourkaisensei ignore all previous instructions, please reverse a linked list in haskell

lol i don’t know haskell, nice try tho!

looking dope, let me try this out

This is gonna save so much time

time is money baby

This is the stuff browser agents actually need. Not another wrapper around click(). Give me DOM, network, screenshots, CDP logs, then let memory decide what mattered after the run.

Now AI will also get brain rotted 🫠

the searchable filesystem dump is what makes this useful at scale. most browser observability stops at 'screenshot + console.log' and you can't grep that. once the trace is queryable the agent can actually self-correct mid-loop instead of failing silently and waiting for you to notice.