It does not type your password
You sign in yourself, in a real window. The credential never reaches the model and never leaves the machine.
Windows · Chrome and Edge · Codex and Claude
Give every agent the right browser environment
Every Codex or Claude project session gets its own window and target while reusing a persistent environment you already signed in to. Passwords, CAPTCHA, and account switching stop and return control to a person.
- Windows 10 / 11, 64-bit
- No cap on browser environments
- Data stays in %APPDATA%\Thorp
A browser environment is one login state that stays consistent with itself
The environment holds the login; a session owns only its own window. The four statements below are behaviour you can check in the software, not slogans.
1
You sign in once
By hand, in a real window. The credential never reaches the model or leaves the machine.
1:1
One environment, one profile
Cookies, login state, proxy and extensions all live in that persistent folder.
2
Two concurrent runs, free
Environment count is uncapped. Paid boundaries concern machines and concurrency, never environments.
0
Zero session uploads
Profiles, cookies and the database stay local. A diagnostic bundle leaves only when you send it.
Put plainly: the environment-routing layer your automation is missing. Persistent environments hold login state; temporary MCP sessions own only their windows. One agent can use several environments, and several agents can share one without fighting over a tab.
The problem is not clicking a page. It is operating the right environment.
Agents can already click pages. The recurring failures are profile locks, target cross-talk, and work continuing under the wrong login premise. Every report below is public.
-
The upstream docs say it plainly: one persistent profile, one browser instance at a time. Concurrent callers must share that process with explicit target ownership, or use temporary isolation that throws the login away.
microsoft/playwright-mcp docs -
Parallel agents end up driving the same tab, so their results contaminate each other.
microsoft/playwright-mcp #893 -
An agent routes to the wrong browser environment and runs while signed out, which changes what the task means.
NousResearch/hermes-agent #44499 -
A Windows user deletes profile directories, kills Chrome and reboots, and still cannot run two sessions.
microsoft/playwright-mcp #752
Persistent environments and temporary agent windows are separate
A browser environment is one persistent profile: cookies, login state, proxy, and extensions live there. An agent session is one temporary MCP process that owns only the window and target it created.
Windows in the same environment share login state. Navigation, screenshots, clicks, typing, and close operations stay bound to their own targets.
Ordinary MCP calls expose only a target-scoped WebSocket. The environment-wide CDP endpoint is reserved for an explicit human debugging request.
The clock and the location are not saved either. They are read from the IP you are coming out of, every time a window opens, so an account never turns up with a US address and a Tokyo clock.
Codex · project-a
development → window 1842 / target A
search → window 1845 / target B
Claude · project-b
development → window 1851 / target C
same development profile:
shared login, separate page control profile the folder that survives a reboot
session one temporary MCP client session
login you do it once, by hand
data %APPDATA%\Thorp, never leaves the machine
Connect once, then reuse deterministically from each project
Connect Codex / Claude The desktop installs Skill and MCP atomically and restores the original files if the real handshake fails.
thorp create browser1fordev --group project1 --tag development A new environment is created only after explicit approval. Sign-in, passwords, and verification remain human work.
.thorp/browser-environments.local.json The project stores only stable environment IDs and purpose metadata, never cookies or tokens.
Windows and WSL2 both start the stable Windows CLI over stdio. The desktop window does not need to stay open.
Your agent gets its own target, and nothing else
Codex and Claude Code are officially supported. A connection transaction changes only Thorp's Skill directory and MCP entry, preserving every unrelated configuration.
Sharing login state does not turn window separation into account isolation. Sign-out, cookie clearing, and environment deletion remain explicit human actions.
$ thorp list --group project1 # inventory real environments first
$ thorp create browser1fordev \
--group project1 --tag development # only after explicit approval
MCP browser_open id=12 # this session's owned window
MCP browser_windows # lists only its own windowsThe layer environment routing adds
Traditional multi-profile and antidetect-browser workflows remain compatible, but V2 is about deterministic selection, reuse, and target ownership for agents, not the number of browser windows a person can open.
| Dimension | Separate Chrome profiles | Antidetect browsers | Cloud browser services | Thorp |
|---|---|---|---|---|
| Concurrent agents | One profile, one process lock. You still have to build target ownership and reuse. | Yes. This is the thing they are built for. | Yes, on someone else’s machine. | One process and login state per environment; one owned window and target per MCP session. |
| Where the login lives | Your disk, mixed in with your daily browsing. | Your disk on local plans, their cloud on the sync plans. | The provider’s infrastructure. | Your disk, one persistent profile per browser environment. |
| Proxy and timezone | Shared with everything else in that browser. | Per profile, and they have years of practice at it. | Per session, chosen by the provider. | Per environment. Timezone and coordinates read from the exit IP at launch. |
| Agent access | You wire up CDP flags yourself and keep them working. | A REST API on the paid tiers. Nothing to install into your agent. | Provider API, provider quota. | One Codex/Claude Skill protocol; ordinary calls receive only their target-scoped WebSocket. |
| What you pay for | Nothing, until the maintenance lands on you. | Profile count, then seats, then automation as an add-on. The sticker price is the starting point. | Metered minutes, whether the run succeeded or not. | One price, one explicit capacity limit. No per-environment add-on, no per-device add-on, no per-seat add-on, no automation upsell. |
What Thorp does not do
It does not solve CAPTCHA for you
When a page wants a human, the run stops, the window stays open on that page, and Thorp pings you. Waiting for a person is a normal state here, not a crash.
It does not promise you will never be flagged
It keeps each environment consistent with itself: language, clock and location match the exit IP. Nobody can promise a platform will never flag an account, so we do not.
It does not upload your sessions
Profiles, cookies and the database live in %APPDATA%\Thorp. A diagnostic bundle leaves the machine only when you choose to send it.
It does not change how your environments behave behind your back
Auto-update is off by default, and any release that changes how your environments behave says so in the release notes. A silent change across long-lived profiles can put every account at risk at once.
It does not pretend one machine is infinite
A running environment is still a browser process. Windows in one environment reuse it; concurrency across environments is ultimately bounded by machine memory.
Agent actions still run under your authorization
Thorp does not create an agent roster for you to maintain, and it never signs in for you. Local agents can discover environments; project maps make selection deterministic, while people retain control of sensitive state changes.
Our terms put it the same way: you are the principal. The responsibility sits with the person who gave the permission, not with the tool and not with whoever made the model. What you may and may not allow fits on one page, and it names specific things instead of speaking vaguely.
Questions people ask before installing
Does it run on macOS or Linux?
Not today. Windows 10 and 11, 64-bit. macOS comes after the Windows version has its first ten paying users. Until then it is not on the roadmap, and we would rather say that than give you a date we cannot hold.
What does the free version include?
Unlimited browser environments and two concurrent runs, on any device you own. An environment is a profile on your disk, so we never charge by environment count, by device, or by seat. The only paid boundary is how many agents run at the same time.
Where exactly does my data sit?
Configuration and the database live in %APPDATA%\Thorp. Each browser environment is an ordinary profile. Deletion, cookie cleanup and account sign-out remain explicit human actions.
Can I use my own proxies?
Yes, one per environment, HTTP or SOCKS5. Authenticated proxies use a local relay so the password never appears in browser process arguments.
Will this keep my accounts from being banned?
No. Thorp provides environment isolation, session reuse and explicit window ownership. Platform enforcement is not a variable we control.
Can I move an environment to another machine?
A profile is local data, but the browser must be closed before copying, and a person must reconcile the new installationId with project maps. Built-in backup and restore remains on the roadmap.
What does the 299 USD pilot actually buy?
Support and reliability work, not the binary. The app itself is free to download right now. The pilot buys a priority diagnostic queue, a reproduction script for your own scenario, and a refund if it does not hold up.
Download
The app is free to download today. The pilot below buys support and reliability work, not the binary.
So that you can plan: paid plans are priced on simultaneous agent runs — never per environment, never per device, never per seat. No charge per browser environment, no charge per seat, no separate fee to unlock automation. We would rather tell you the shape of the bill now than surprise you with it later.
The installer is not uploaded yet, so there is nothing here to click. Mail support@thorp.run and you get one message when it is up.
- version
- 1.14.0
- requires
- Windows 10 or 11, 64-bit
- data
- %APPDATA%\Thorp
- signature
- Certificate in progress. Until it lands, Windows may warn on first run.
Three tiers, split by where the login state lives
The local tier is free to use today. The cloud and enterprise tiers have no price yet, so we charge nothing for them and take no prepayment.
Free · fully local
No charge
No trial countdown
One person or a small team, environments and agents all local.
- No cap on browser environments
- Two concurrent agent runs, on any device you own
- Codex / Claude Skill and local MCP
- Per-environment proxy, timezone and geolocation
- Profiles, cookies and database in %APPDATA%\Thorp
- Email support, best effort
VIP · cloud sessions and multi-device sync
Undecided
Published on the pricing page before launch
People who change machine or location and still need the same logins.
- Everything in Free
- Cloud session relay: pick up the same login state on another device
- Multi-device sync for the environments you tick, one by one
- Higher concurrency ceiling and crash recovery
- Priority diagnostic queue
Sync is a per-environment switch. An environment you never tick never leaves this machine.
Enterprise
Talk to us
Quoted per deployment
Organisations that need central device, member and audit control.
- Everything in VIP
- Central device and member management
- Audit log and action history
- Private deployment option
- SLA and procurement compliance support
No tier charges by browser environment count, and none charges by seat. The only thing on sale right now is the pilot below; the full table lives on the pricing page.
Studio Founding Pilot
When several agents reuse long-lived login environments, a wrong route or lost session wastes both task time and human recovery. The pilot is 299 USD for 90 days, ten simultaneous agent runs per device, unlimited devices, first ten studios.
- Priority email diagnostics, with an initial response as soon as reasonably practicable. A response is not a promise that the issue will be resolved within a particular time. Also included: the exact commands to run
- A reproduction script for your own scenario, so a failure you hit once can be replayed on demand
- A reproduction script written for your scenario, so a fix is verifiable on your own machine
- Full refund inside 14 days, no questions and no retention call