1. Not detectable and so not blocked
2. Uses my IP (I have separate hardware available)
3. Uses cookies from my main browsing sessions
4. Run in parallel-shadow so it doesn’t block me,
I don’t break it.
We are in the cat and mouse world of scraping, but with end users now. Companies that enable agents will thrive and those that don’t, won’t.
I also wonder how much of a SMB profit margin comes from upsells and dark patterns. Agents help me a lot with avoiding those, also reducing my cognitive complexity. I’m at the point where I don’t order things if I can’t delegate my agent, I don’t want to see all their bullshit that they’ve structured to extract more money from me. My agents save me from those costs.
My setup covers your first three, though not with a dedicated agent browser. The agent (Claude Code) drives an ordinary Chrome that I use myself, through the Chrome DevTools Protocol (CDP). That Chrome is started with its own profile and Chrome's remote-debugging switch, and I'm thus automatically logged into Gmail, Slack, shops etc. So it's my IP, my cookies, and the browser that shops don't block. The agent opens a temporary tab, reads what it needs, and closes it - unless it needs to leave it open for me to actually make the purchase.
It's one browser, so the agent and I share it, so it doesn't pass your fourth point.
Seems like we share the same frustrations with most dedicated agentic browsers. I maintain an firefox fork that does all of this (cookie syncing you can perform manually via the API when you'd like):
Out of curiosity, what prompted you to make this when agent-browser.dev and lots of other alternatives already exist? "Because it's fun" is absolutely a legitimate answer, but I couldn't tell from the README what I might be missing.
It can, but it can also use your personal browser (Chrome and Safari, haven’t tried Firefox).
> Chrome for Testing is the default browser. You can also attach to a running Chrome with --cdp or --auto-connect (CDP mode), select the Lightpanda or Obscura engine with --engine, drive Mobile Safari in the iOS Simulator, or use a cloud browser provider such as Browserbase, Browserless, Browser Use, Kernel, or AgentCore.
I love it! I’m going to give this a try for sessions that need to do a lot of browsing. If anyone is looking for a headless WebKit browser/viewer for agents (more focused on development), check out https://github.com/bensyverson/sleepyhollow
Great work, I'll try it out as a replacement for chrome for my verification flows today. It's great to have ad blocker but it's curious to me to enable it by default, is this something you use more for scraping than application testing for example?
With ad blocker it is less context for ai, faster and less ram usage (rules size is less than some ad blockes). I use it to automate things in work and personal life, debugging, testing frontend, running auto tests.
Nice! Most of the time I'm doing tests on my own web applications rather than on the open web and focus hijacking during test cycles was becoming a real pain. Plus I resent Chrom*'s outsize footprint.
We use agweb daily and love it. It is a good alternative to Atlas that gives you a full desktop (most of our use cases are web, of course), and it doesn't require a new user.
What's your specific usecase? Just general browser as a service?
There are a ton: kernel, browserbase, browser-use.
Any more requirements other than cloud hosted?
Thanks for chiming in! It needs to be self-hosted/open source, and a platform that can be used by any agent (via OpenCode) as well as through Hermes Bot Screen. I need to also be able to take over the browser/screen to login, fill out captchas, etc (just like Grok Bot). Does that help?
Are the UBlock Origin lists configurable and is the browser stealthy enough to been indistinguishable from normal chrome when used from a residential IP?
Lists only via env for now, AGWEB_ADBLOCK_LISTS. It's real WebKit and says Safari, not Chrome. Clicks and typing are native events at human pace, no fingerprint spoofing though.
> agweb has no automatic cookie-banner remover. Its ad blocker uses uBlock Origin’s lists, which block ads and trackers but leave consent dialogs alone
I mean, rather than just out-and-out removing them, I would think that the proper thing to do to get those dialogs out of the way would be to have a separate extension that detects them and automatically interacts with them to dismiss them (in whatever way you'd prefer.)
I run several agents at once. Playwright's Chrome windows kept popping up and grabbing my keyboard, and headless got blocked by some sites and you can't see easily what's happneing.
agweb keeps what's good about headless and still lets you watch. It never pops up or takes focus, and idle tabs get suspended: five idle pages sit at 0% CPU and 590 MB, vs 1.6 GB for Playwright's Chrome. ⌘D shows every agent's tabs live.
For agents, the commands mirror playwright-cli, which they already know. Each reply also says what the action changed (fetches, elements added or removed), so they need fewer snapshots, fewer tool calls and fewer tokens.
Also: one login shared by all agents through a named profile, fill --secret-cmd 'pass show site' keeps passwords out of the transcript, and ask hands the tab to you for MFA or a CAPTCHA.
The shared named profile is handy, but how do you deal with one agent's actions spilling over to the others? For example, one agent logs out, switches accounts, or trips a site's rate limit, and every other agent on that profile inherits it. Is there a way to give an agent a forked or read-only copy of the profile's cookies?
This - for some scenarios I want the cookies shared (multiple agents watching eg fb marketplace), for others I want to have their own sandbox. And the GUI could indicate this eg. by putting tabs that share profile into blocks.
The big features I need in an agentic browser.
1. Not detectable and so not blocked 2. Uses my IP (I have separate hardware available) 3. Uses cookies from my main browsing sessions 4. Run in parallel-shadow so it doesn’t block me, I don’t break it.
We are in the cat and mouse world of scraping, but with end users now. Companies that enable agents will thrive and those that don’t, won’t.
I also wonder how much of a SMB profit margin comes from upsells and dark patterns. Agents help me a lot with avoiding those, also reducing my cognitive complexity. I’m at the point where I don’t order things if I can’t delegate my agent, I don’t want to see all their bullshit that they’ve structured to extract more money from me. My agents save me from those costs.
My setup covers your first three, though not with a dedicated agent browser. The agent (Claude Code) drives an ordinary Chrome that I use myself, through the Chrome DevTools Protocol (CDP). That Chrome is started with its own profile and Chrome's remote-debugging switch, and I'm thus automatically logged into Gmail, Slack, shops etc. So it's my IP, my cookies, and the browser that shops don't block. The agent opens a temporary tab, reads what it needs, and closes it - unless it needs to leave it open for me to actually make the purchase.
It's one browser, so the agent and I share it, so it doesn't pass your fourth point.
Seems like we share the same frustrations with most dedicated agentic browsers. I maintain an firefox fork that does all of this (cookie syncing you can perform manually via the API when you'd like):
https://github.com/MonkeySee-AI/rotunda
Linking if helpful.
Out of curiosity, what prompted you to make this when agent-browser.dev and lots of other alternatives already exist? "Because it's fun" is absolutely a legitimate answer, but I couldn't tell from the README what I might be missing.
"because it's fun" is a highly underutilized reason
What alternative show you a visual preview like this?
https://github.com/anomalyco/browse http://browsermcp.io/ and others, no doubt.
The features I look for a persistent sessions, so I can login, visual feedback tools, and low resource usage.
I assume all can do this for human-in-the-loop scenarios. I've personally settled on agent-browser.dev for now.
agent-browser uses chrome test mode or whatever it is called, which means a bunch of stuff gets broken by bot detection.
It can, but it can also use your personal browser (Chrome and Safari, haven’t tried Firefox).
> Chrome for Testing is the default browser. You can also attach to a running Chrome with --cdp or --auto-connect (CDP mode), select the Lightpanda or Obscura engine with --engine, drive Mobile Safari in the iOS Simulator, or use a cloud browser provider such as Browserbase, Browserless, Browser Use, Kernel, or AgentCore.
I love it! I’m going to give this a try for sessions that need to do a lot of browsing. If anyone is looking for a headless WebKit browser/viewer for agents (more focused on development), check out https://github.com/bensyverson/sleepyhollow
Great work, I'll try it out as a replacement for chrome for my verification flows today. It's great to have ad blocker but it's curious to me to enable it by default, is this something you use more for scraping than application testing for example?
With ad blocker it is less context for ai, faster and less ram usage (rules size is less than some ad blockes). I use it to automate things in work and personal life, debugging, testing frontend, running auto tests.
How are you planning to handle multiple agent sessions running concurrently? Agents can jump into other session! possibilities are high
Nice! Most of the time I'm doing tests on my own web applications rather than on the open web and focus hijacking during test cycles was becoming a real pain. Plus I resent Chrom*'s outsize footprint.
We use agweb daily and love it. It is a good alternative to Atlas that gives you a full desktop (most of our use cases are web, of course), and it doesn't require a new user.
I’m looking for something similar to this but can be self hosted in the cloud. Any suggestions?
What's your specific usecase? Just general browser as a service? There are a ton: kernel, browserbase, browser-use. Any more requirements other than cloud hosted?
Thanks for chiming in! It needs to be self-hosted/open source, and a platform that can be used by any agent (via OpenCode) as well as through Hermes Bot Screen. I need to also be able to take over the browser/screen to login, fill out captchas, etc (just like Grok Bot). Does that help?
Are the UBlock Origin lists configurable and is the browser stealthy enough to been indistinguishable from normal chrome when used from a residential IP?
Lists only via env for now, AGWEB_ADBLOCK_LISTS. It's real WebKit and says Safari, not Chrome. Clicks and typing are native events at human pace, no fingerprint spoofing though.
> agweb has no automatic cookie-banner remover. Its ad blocker uses uBlock Origin’s lists, which block ads and trackers but leave consent dialogs alone
this is awesome
What's good about cookie consent dialogs?
A lot of pages don't work when an adblocker removes the Dialog.
I mean, rather than just out-and-out removing them, I would think that the proper thing to do to get those dialogs out of the way would be to have a separate extension that detects them and automatically interacts with them to dismiss them (in whatever way you'd prefer.)
I built something similar (albeit didn't put too much effort in), so will be moving to this! Great work.
Thanks!
Looks interesting! But..
> Needs macOS 26 or later.
Too bad. I'll check back when someone submits a PR for Linux.
PR is welcome, happy to support it!
is there a cli version / headless or build on such.
I run several agents at once. Playwright's Chrome windows kept popping up and grabbing my keyboard, and headless got blocked by some sites and you can't see easily what's happneing.
agweb keeps what's good about headless and still lets you watch. It never pops up or takes focus, and idle tabs get suspended: five idle pages sit at 0% CPU and 590 MB, vs 1.6 GB for Playwright's Chrome. ⌘D shows every agent's tabs live.
For agents, the commands mirror playwright-cli, which they already know. Each reply also says what the action changed (fetches, elements added or removed), so they need fewer snapshots, fewer tool calls and fewer tokens.
Also: one login shared by all agents through a named profile, fill --secret-cmd 'pass show site' keeps passwords out of the transcript, and ask hands the tab to you for MFA or a CAPTCHA.
macOS 26+, WebKit only, no CDP.
Why would you mirror the playwright CLI but not CDP? This looks pretty cool, but it's completely useless to me without CDP.
Probably because it’s not chromium
agweb is the system WebKit, so no CDP. Commands copy playwright-cli bc agents already know them. If you need CDP, Playwright fits better.
The shared named profile is handy, but how do you deal with one agent's actions spilling over to the others? For example, one agent logs out, switches accounts, or trips a site's rate limit, and every other agent on that profile inherits it. Is there a way to give an agent a forked or read-only copy of the profile's cookies?
This - for some scenarios I want the cookies shared (multiple agents watching eg fb marketplace), for others I want to have their own sandbox. And the GUI could indicate this eg. by putting tabs that share profile into blocks.
Not yet. For now, state-save in one workspace and state-load in a fresh one gives an in-memory copy, so a logout stays there. Fork flag can be done.