CloakBrowser Agent: a Jev-powered stealth browser agent
Give it a goal in plain language. TypeSafe Jev decides every step in ~0.3 s, CloakBrowser carries it out like a human, and you get the result back as markdown.
Agenten stärken
Installation vom Herausgeber. Kein PlugMCP-Endpunkt.
Host a copy with our keys and logs.
Noch kein paketiertes Release in der Registry. Folgen Sie der README des Repositorys.
Remote-URL oder ein Befehl auf Ihrem Rechner? Worin sich die beiden Installationen unterscheiden
Auf PlugMCP ausführen
A new server you create. It is not this listing's official endpoint. Use the publisher install when you want their URL and their credentials.
StartenVom Herausgeber
Give it a goal in plain language. TypeSafe Jev decides every step in ~0.3 s, CloakBrowser carries it out like a human, and you get the result back as markdown.
Jev-powered stealth browser agent. TypeSafe Jev decides each step in ~0.3 s, CloakBrowser carries it out like a human. MCP server, CLI and Python API.
browse(goal="Find one-way flights from Zurich to London on October 20, 2026, for one adult in economy.
Stop when matching flight options are visible.",
url="https://www.google.com/travel/flights?hl=en")
status: done
tab_id: t1 (still open)
url: https://www.google.com/travel/flights/search?tfs=...
title: Zürich to London | Google Flights
steps: 10 actions, 14 decisions, 22596 ms
actions taken (p = Jev's probability for the chosen target; runner-ups in brackets):
1. click 'Change ticket type. Round trip' p=0.81 ['Open Where from?' p=0.15, 'Flights' p=0.01]
2. click 'One way' p=1.0
3. fill 'Where from?' = 'Zurich' p=0.93 ['Where to? ' p=0.04, 'Departure' p=0.03]
4. click 'Zürich, Switzerland' p=0.77 ['Zurich Airport (ZRH)' p=0.21, 'Open Where from? ' p=0.01]
5. fill 'Where to? ' = 'London' p=0.99 ['Departure' p=0.01]
6. click 'London, United Kingdom' p=0.95 ['Heathrow Airport (LHR)' p=0.02, ...]
7. click 'Open Departure' p=0.97 ['Search' p=0.03]
8. click 'Tuesday, October 20, 2026 , 48 euros, Cheapest price' p=0.98 ['Done. ' p=0.02]
9. click 'Done. Search for one-way flights, departing on October 20, 2' p=0.98 [...]
10. click 'Search' p=0.99 ['Open Where from?' p=0.01]
...
<untrusted_page_content>
# Flight search
...
## Search results
...
from €48
A real run, trimmed.
It works as an MCP server (Claude Code, Cursor, Claude Desktop, any MCP client), a CLI, or a Python library. It runs on CloakBrowser, a stealth Chromium with human-like mouse and keyboard input.
Most browser agents ask a large language model to write the next action, which costs seconds per step. Jev is TypeSafe's first System One model. It doesn't generate text: you give it the current state plus typed questions, and it returns an answer with calibrated probabilities. A browser step is exactly that kind of question: which of these controls, doing what?
goal ─► OBSERVE read the page → numbered table of the controls a user can actually reach
▲
│ DECIDE one Jev request: which operation (CLICK, TYPE_TEXT, SELECT, SCROLL, WAIT, DONE, BLOCKED)
│ and which element, answered together
│ TYPE_TEXT → a small text model writes only the value for that one field
│ (or a caller secret is typed locally, see Logins)
│
└─ ACT human-like click / typing, after checking the page did not change
DONE → a second check asks Jev "achieved?" and "NOT achieved?" about the page (disagreement → done_unverified)
→ the page is turned into markdown, Jev scores each section against the goal, the best sections are returned
blocked, as do three actions in a row that change nothing.cloakbrowser login or go to cloakbrowser.dev/free. A free key allows one browser session at a time, and a paid key raises that limit. Without any key, the older build is used.The browser binary downloads automatically on first use. Node is not needed.
pip install cloakbrowser-agent
This installs the cloak-agent and cloak-agent-mcp commands, plus cloakbrowser (for cloakbrowser login).
For MCP clients you don't even need to install it: the configs below use uvx, which fetches and runs it on demand.
From source: git clone https://github.com/CloakHQ/CloakBrowser-Agent && cd CloakBrowser-Agent && pip install -e .
Claude Code
claude mcp add cloak-agent --scope user \
-e TYPESAFE_API_KEY=... \
-e TEXT_MODEL_BASE_URL=https://openrouter.ai/api/v1 -e TEXT_MODEL=... -e TEXT_MODEL_API_KEY=... \
-e CLOAKBROWSER_LICENSE_KEY=cb_... \
-- uvx --from cloakbrowser-agent cloak-agent-mcp
Cursor / Claude Desktop (mcpServers in the client's config)
{
"mcpServers": {
"cloak-agent": {
"command": "uvx",
"args": ["--from", "cloakbrowser-agent", "cloak-agent-mcp"],
"env": {
"TYPESAFE_API_KEY": "...",
"TEXT_MODEL_BASE_URL": "https://openrouter.ai/api/v1",
"TEXT_MODEL": "...",
"TEXT_MODEL_API_KEY": "...",
"CLOAKBROWSER_LICENSE_KEY": "cb_..."
}
}
}
}
browse(goal, url?, tab_id?, secrets?) runs one whole task and returns:
status: done, done_unverified (the agent stopped as done, but a second check of the page disagreed: look before trusting it), blocked (no control can make progress, or the same action came up on the same page a 3rd time), needs_input (the goal lacks a value a field needs), budget (step limit reached), or errortab_id (the tab stays open), the final URL and title, and the number of actions, decisions and millisecondsp for it, the top runner-ups in brackets, and op=: its probability for the operation itself, with the runner-up operation (op=CLICK 0.43 (SCROLL_DOWN 0.35)). A low p or op, or a runner-up close behind, shows where the agent was unsure. In our tests, wrong clicks showed up in a low op more than in p.done): the straight and negated "is the goal achieved?" probabilities<untrusted_page_content>While a call runs, each step is also sent as a live progress notification (clients that show MCP progress display it).
Logins without exposing the password: pass secrets={"username": "env:SHOP_USER", "password": "env:SHOP_PW"} and set those variables in the MCP server's env. Only env:NAME references are accepted, so the values never pass through the calling model. Jev sees only the names and picks which one fits each field. The value is typed locally and appears everywhere (steps, page content, every model request) as [secret:name]. A password field with no matching secret ends as needs_input. The text model never writes a password.
snapshot(tab_id, screenshot?) shows a tab exactly as the agent sees it: URL, title, scroll position, the operations on offer, the numbered element table ([3] button "Reject all") and the visible text. With screenshot=True it also returns an image of the visible part of the page. Use it to see why a task stopped.
act(tab_id, op?, target?, instruction?, text?, screenshot?) does one step in a tab yourself, for example to get past a page the agent is stuck on. Give either:
op + target from the latest snapshot: CLICK "3", TYPE_TEXT "5" with text, SELECT "4:2", or SCROLL_DOWN / SCROLL_UP / WAIT. No model is involved: text is typed exactly as given. Password fields show a masked value. To fill one without the value passing through you, use browse with secrets.instruction, one step in plain language ("click Reject all"). Jev picks the element; the reply shows its p and runner-ups.act returns what it did, whether the page changed, and the new snapshot, so steps can be chained. If the page changed since the snapshot, nothing is done and it answers status: stale. A tab serves one call at a time; a second call on it answers busy. Continue the task with browse(goal, tab_id=...) whenever you like.
close_tab(tab_id) closes a tab.
A tab_id stays valid until you close_tab it, close it in the browser, or the browser idles out. After that, browse answers status: error (tab 't1' is gone; open tabs: ...).
# 1. A task that needs values the goal did not include
browse(goal="Fill in the pizza order form and submit it.", url="https://httpbin.org/forms/post")
→ status: needs_input (No value in the goal for field: Customer name:) tab_id: t1
browse(goal="Fill in the pizza order form with customer name Jane Doe, telephone 555-0100, "
"email [email protected], size medium, and submit it.", tab_id="t1")
→ status: done (fills all four fields on the same form, clicks 'Submit order')
# 2. A follow-up step on the page the last task ended on
browse(goal="Search Google for 'CloakBrowser GitHub' and open the CloakHQ/CloakBrowser repository.",
url="https://www.google.com")
→ status: done tab_id: t2
browse(goal="Open the Issues tab of this repository and list the titles of the newest issues.", tab_id="t2")
→ status: done (1 action: click 'Issues')
# 3. Taking over a stuck task: look, do one step by hand, let the agent continue
snapshot(tab_id="t3")
→ [2] searchbox "Search Wikipedia" ops=TYPE_TEXT,CLICK ...
act(tab_id="t3", op="TYPE_TEXT", target="2", text="Kurt Gödel")
→ status: done (new snapshot: [3] option "Kurt Gödel Mathematician and philosopher (1906–1978)" ...)
act(tab_id="t3", op="CLICK", target="3")
→ status: done url: https://en.wikipedia.org/wiki/Kurt_G%C3%B6del
browse(goal="Summarize his incompleteness theorems.", tab_id="t3")
Tips:
browse calls can run at once. Each gets its own tab in the same browser.CLOAK_AGENT_IDLE_MINUTES without calls (default 5), the browser closes itself along with its tabs, and it relaunches on the next call. It also closes when the MCP server stops.A profile can only be open in one browser at a time. Give each instance its own CLOAK_AGENT_PROFILE.
One identity per profile. Each profile keeps its own fingerprint seed across launches, so a returning visitor (same cookies, same IP) looks like the same device every time. If you switch CLOAK_AGENT_PROXY to a different exit or country, switch to a new CLOAK_AGENT_PROFILE too. That way IP, cookies and device change together.
The exit IP also decides which regional version of a site you get. If a store isn't available in your location, use a proxy in a country where it is.
export TYPESAFE_API_KEY=... TEXT_MODEL_BASE_URL=https://openrouter.ai/api/v1 TEXT_MODEL=... TEXT_MODEL_API_KEY=...
export CLOAKBROWSER_LICENSE_KEY=cb_... # or run `cloakbrowser login` once
# one task, own browser (closes at the end; add --keep-open to inspect)
cloak-agent run --url https://en.wikipedia.org --goal "Open the Wikipedia article about Gödel's incompleteness theorems."
# debugging: keep one headed browser running, then run tasks against it
cloak-agent browser --port 9222
cloak-agent run --cdp http://127.0.0.1:9222 --url https://www.google.com --goal "Search Google for 'CloakBrowser'"
--no-humanize switches to instant input.
Uses the same environment variables as above, including CLOAKBROWSER_LICENSE_KEY (or the key saved by cloakbrowser login).
import asyncio
from cloak_agent import Session, run
async def main():
session = await Session.launch(headless=False) # or: await Session.connect("http://127.0.0.1:9222")
page = await session.new_tab("https://en.wikipedia.org")
result = await run(session, "Open the Wikipedia article about Gödel's incompleteness theorems.", page=page)
# login: run(session, "Log in", page=page, secrets={"username": "...", "password": "..."})
print(result["status"], result["url"])
print(result["markdown"])
await session.close()
asyncio.run(main())
Most of a humanized run is spent moving the mouse and typing at a human pace. That's on purpose, because behavior is scored on protected sites.
Jev decisions take ~0.3 s each, roughly 1 s per task. Filling a field with the text model adds 0.5–2 s depending on the model.
<select> values, and date/time fields (date, time, datetime-local, month, week), are set programmatically, not through a mouse-driven dropdown or picker. Sliders (range) and color pickers aren't offered.secrets.done means the model saw the goal as met and a second check agreed. Check results that matter.<untrusted_page_content> so the calling agent doesn't treat it as instructions.The observe → choose → act loop, the element snapshot and the Jev instructions are adapted from browser-use/jev-ultrafast (MIT, see LICENSE-jev-ultrafast).
Jev-powered stealth browser agent. TypeSafe Jev decides each step in ~0.3 s, CloakBrowser carries it out like a human. MCP server, CLI and Python API.
Tools
Noch keine Tool-Liste. Remote-Server erhalten eine nach dem ersten Health-Check.
FAQ
Kopieren Sie auf dieser Seite die stdio-Konfiguration für Cursor, Claude Desktop, Claude Code oder VS Code. Sie nutzt die Installation des Herausgebers, keinen PlugMCP-Proxy.
In diesem Eintrag ist keine erforderliche Umgebungsvariable aufgeführt.
CloakBrowser Agent läuft lokal über stdio.
Zuletzt geprüft: 2026-11-08. Jev-powered stealth browser agent. TypeSafe Jev decides each step in ~0.3 s, CloakBrowser carries it out like a human. MCP server, CLI and Python API.