Wraith MCP Server
Provides an MCP-native agent browser that enables autonomous agents to perceive and interact with web pages through stealth browsing, identity borrowing, and WAAP detection.
README
🌀 Wraith
A stealth, identity-borrowing, MCP-native agent browser.
Wraith is a Python toolkit that gives an autonomous agent a real browser that is hard to fingerprint and easy to drive. It pairs a hardened Firefox engine (Camoufox) with identity borrowing — reusing a warmed, already-authenticated session from one of your own real browser profiles — and exposes the whole thing as both a CLI and an MCP server so an LLM can perceive and act on pages by index.
⚖️ Responsible Use & Legal
Wraith is a dual-use tool. Like the established projects in its category (Camoufox, nodriver, browser-use, undetected-chromedriver), it can be used well or badly. It is published for legitimate purposes:
- Accessing your own accounts and data, on your own machine, with your own warmed browser profile.
- Authorized security testing and bot-defense research (where you have permission to test the target).
- Personal automation, scraping of data you are entitled to, and reproducible research into anti-bot / WAAP systems.
Please use it responsibly:
- Respect each target site's Terms of Service and all applicable laws (e.g. CFAA and its equivalents).
- Identity borrowing reads only your own local browser profiles. Do not use it against accounts or sessions that are not yours.
- Do not use Wraith for fraud, account takeover, credential stuffing, mass abuse, spam, or to circumvent access controls you are not authorized to bypass.
You are responsible for how you use this software. If your use isn't clearly covered by the legitimate cases above, don't do it. Wraith is released under the MIT License with no warranty.
Why identity borrowing (not "solving")
Modern defenses like reCAPTCHA-v3 and Reblaze/Link11 ac_v2 are
reputation systems, not puzzles. reCAPTCHA-v3 has no client-side solver: it
returns a 0.0..1.0 score derived from your Google-account cookies, aged
browsing history, and IP reputation. A fresh automated profile scores ~0.1–0.3
("bot") no matter how good the stealth engine is, because it has no history to
vouch for it. You cannot fake reputation.
So Wraith doesn't try. For your own sites and accounts, it borrows a warmed identity: it reads your live session cookies straight out of your real Firefox or Zen profile on disk (or harvests a live bearer token from network traffic) and injects them into the stealth context. The agent then navigates as the already-authenticated user — the reputation comes along for free.
For sites with a solvable JS interstitial (e.g. Reblaze ac_v2), no identity
is needed at all: a real Firefox engine clears the challenge natively, and
clear_challenge() just waits for the clearance cookie to appear.
Features
- Stealth-first engine — Camoufox (hardened Firefox) primary, patchright
Chromium fallback; enforces the
playwright==1.55pin Camoufox needs. - Identity borrowing — discover and read cookies from your own Firefox / Zen / Chrome profiles and inject them into the stealth context.
- Live session harvesting — latch the first request carrying both an
Authorizationheader and a named auth cookie, for tokens never stored on disk. - WAAP/bot-defense fingerprinting —
identify_waap()recognises Reblaze/ Link11, Akamai, reCAPTCHA, DataDome, Imperva/Incapsula, Kasada, SiteMinder. - Self-assessment — read your own reCAPTCHA-v3 score and rebrowser automation tells.
- Agent perception layer — browser-use-style indexed DOM snapshots; act on
elements by index (
click,type,scroll,read). - MCP-native — a stdio MCP server (
wraith-mcp) exposing the agent browser as tools for any MCP client. - Residential proxy support — a dependency-free
ProxyPoolplus a first-class DataImpulse provider for rotating/sticky exits. - Human-like behavior helpers — curved/eased mouse movement and per-key typing cadence.
- Resilient import — a missing optional browser dep never breaks
import wraith; it's recorded inwraith.missing_imports.
Install
Wraith uses uv:
uv sync # core deps (camoufox, playwright==1.55, patchright, httpx, mcp)
uv run camoufox fetch # fetch the Camoufox Firefox build (primary engine)
uv run patchright install chromium # (optional) fetch patched Chromium for the fallback engine
Or with pip:
pip install wraith
camoufox fetch # fetch the Camoufox Firefox build
patchright install chromium # (optional) fallback engine
Verify the install:
uv run python -c "import wraith; print(wraith.__version__)"
uv run wraith --help
⚠️ Camoufox 0.4.x crashes on
playwright >= 1.60(a FirefoxpageErrorserialization bug). Wraith pinsplaywright == 1.55.xand detects a mismatch up front with an actionable error. patchright is independently versioned and unaffected.
Quickstart
(a) Library — AgentBrowser
The agent browser is the highest-level entry point: navigate, perceive an indexed snapshot, and act on elements by index.
from wraith import agent_browser
# Self-launches a stealth Camoufox session; closes it on exit.
with agent_browser(engine="camoufox", headless=True) as ab:
snap = ab.navigate("https://example.com")
print(snap.to_text()) # [12]<button role=button>Search</button> ...
ab.type(3, "wraith", enter=True) # type into element [3] and press Enter
ab.click(12) # click element [12]
print(ab.read()) # current page as markdown
Lower-level: launch a session and borrow a warmed identity from your own profile.
import wraith
with wraith.browser(engine="camoufox", geoip=True) as s:
# Borrow your own warmed identity from a real Zen/Firefox profile on disk.
profile = wraith.find_zen_profiles()[0]
cookies = wraith.extract_cookies(profile, domain_filter="example.com")
wraith.inject_cookies(s.context, cookies)
s.page.goto("https://example.com") # navigate as the logged-in user
# Optionally harvest a live bearer token the app mints per session.
h = wraith.SessionHarvester(target_url="api.example.com", auth_cookie="session")
h.attach(s.context)
s.page.goto("https://example.com/dashboard")
h.wait(timeout=60)
h.save_session("example.session.json")
import wraith is resilient: if an optional browser dependency is missing, the
affected symbols are omitted and wraith.missing_imports records why.
(b) CLI — wraith
The wraith console script has seven subcommands. The default engine is
camoufox; pass --engine chromium for the patchright fallback or
--engine auto to let Wraith choose.
# agent — open a URL and print a browser-use-style indexed snapshot
uv run wraith agent https://example.com
uv run wraith agent https://example.com --json
# borrow — inject your own warmed cookies, open the site as that logged-in user
uv run wraith borrow https://example.com --host example.com
uv run wraith borrow https://example.com --profile "~/Library/Application Support/zen/Profiles/xxxx.default"
# harvest — capture a live {Authorization, Cookie, User-Agent} session
uv run wraith harvest https://example.com --target api.example.com --cookie session -o example.session.json
# score — read this identity's reCAPTCHA-v3 reputation (fresh -> ~0.1-0.3)
uv run wraith score
uv run wraith score --engine chromium --json
# detect — fingerprint a URL's bot/WAAP defenses
uv run wraith detect https://example.com
uv run wraith detect https://example.com --json
# launch — just open a stealth browser, held open (headed)
uv run wraith launch https://example.com
uv run wraith launch https://example.com --headless --no-wait
# mcp — run the MCP server over stdio (see below)
uv run wraith mcp
Chrome cookies are AES-encrypted via the OS keychain; Wraith deliberately does not decrypt them and raises
ChromeEncryptionErrorwith guidance. Use a Firefox or Zen profile (same engine family as Camoufox, plaintext-readable cookie store) — that's the recommended path — or harvest live.
(c) MCP server
Wraith ships a stdio MCP server (wraith-mcp) that exposes the agent browser as
tools — navigate, snapshot, click, type_text, scroll, read,
screenshot, detect_waap, and borrow. Wire it into an MCP client such as
Claude Code:
claude mcp add wraith -- uv run --directory /path/to/wraith wraith-mcp
(Equivalently, the server starts via uv run wraith mcp or uv run wraith-mcp.)
Architecture
Wraith is a set of focused, mostly-independent modules under wraith/:
| Module | Responsibility |
|---|---|
engine |
Stealth launcher & engine selection. launch() / browser() return a Session (.page, .context, .browser); clear_challenge() is the cookie-free WAAP front door. Camoufox primary, patchright Chromium fallback; enforces the playwright==1.55 pin. |
identity |
Signature feature. Discover Firefox/Zen/Chrome profiles, extract_cookies(), normalize them, and inject_cookies() into a context. Firefox/Zen plaintext; Chrome raises ChromeEncryptionError. |
harvest |
Live session capture. SessionHarvester latches the first request carrying an Authorization header + auth cookie; harvest_session() is the high-level helper. |
detect |
Diagnostics: identify_waap() (vendor fingerprinting), recaptcha_v3_score() (reputation read), bot_detector() (rebrowser automation tells), fingerprint(). |
behavior |
Human-like helpers: human_move() (curved/eased/jittered mouse), human_type() (per-key cadence), dwell(). |
agent |
The perceive/act-by-index browser wrapper. AgentBrowser / agent_browser() built on the snapshot layer. |
snapshot |
Agent perception: take_snapshot() builds an indexed, browser-use-style DOM Snapshot of interactive Elements. |
recaptcha |
v3 token harvesting from a warmed/borrowed session (harvest_token, score) + solver-service skeletons (SolverService, CapSolver, TwoCaptcha). |
proxy |
Dependency-free ProxyPool (round-robin / random) and normalize_proxy() for clear_challenge rotation. |
providers |
First-class residential-proxy integrations. DataImpulse builds proxy URLs (rotating/sticky) and ProxyPools (pool) for launch(proxy=...) / clear_challenge(proxy_pool=...). |
mcp |
The wraith-mcp FastMCP stdio server exposing the agent browser as MCP tools. |
cli |
The wraith command. Lazily imports each component so --help works on a partial install. |
Runnable, heavily-commented examples live in examples/:
borrow_session.py, score_check.py, detect_waap.py.
How it gets past defenses (honest tiers)
Wraith uses the cheapest mechanism that works for each tier — there is no magic universal bypass:
- Engine stealth (Camoufox / Firefox). Firefox skips the Chrome-specific
isChrome()detection cluster entirely, so the bulk of fingerprint tells never run. This alone clears solvable JS interstitials like Reblaze/Link11ac_v2—clear_challenge()just navigates and polls for the clearance cookie. No cookies, no proxy needed. - Identity borrowing. For reputation defenses (reCAPTCHA-v3) and
already-authenticated areas of your own accounts, inject a warmed session
from your real on-disk profile (
extract_cookies→inject_cookies), or harvest a live bearer token (SessionHarvester). You borrow the reputation rather than trying to fake it. - Residential exit rotation. IP-reputation tiers (Reblaze 474/481
rate-limit) can't be cleared by waiting or by cookies — they need a different
exit IP.
ProxyPooland theDataImpulseprovider feed rotating/sticky residential exits intolaunch(proxy=...)andclear_challenge(proxy_pool=...). - Not solvable. Hard blocks (HTTP 492, non-browser /
HeadlessChromeUA) and reCAPTCHA-v3 with no warmed identity are not bypassable by Wraith — they raise actionable errors rather than pretending. reCAPTCHA-v3 has no client solver;recaptcha.pyonly harvests a token from a warmed session.
Docs
- docs/DETECTION.md — a WAAP / bot-system taxonomy with
a quick-ID cheat sheet (header / cookie / status signals) and per-vendor deep
dives (Reblaze/Link11
ac_v2, Akamai, reCAPTCHA-v3, DataDome, Imperva, Kasada, SiteMinder). - docs/PLAYBOOK.md — the decision playbook: engine
choice, the
playwright==1.55pin, Chromium hardening, identity consistency, cookie extraction, live token harvesting, proxy rotation, and an end-to-end decision flow. - docs/AGENTS.md — the agent perception/action layer and the MCP server: snapshot format, acting by index, and wiring Wraith into an MCP client.
Tests
uv run pytest -q
The suite runs fully offline (no browser binaries, no network): it asserts the package imports cleanly, the public API symbols exist, and the detection / identity / proxy logic behaves against synthetic inputs.
Contributing
Contributions are welcome. Please:
- Fork and branch from
main. uv syncand keep the suite green (uv run pytest -q).- Lint with
uv run ruff check .before opening a PR.
Issues and PRs: https://github.com/YogevKr/wraith.
License
MIT — see LICENSE.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
Neon Database
MCP server for interacting with Neon Management API and databases
E2B
Using MCP to run code via e2b.
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.