vibatchium

Agent-piloted browser automation that clears Cloudflare. Patched Playwright + multi-session daemon + credential vault + vision clicking + prompt-injection safety. One MCP server, N parallel Chromes, persistent per-session profiles.

Where this fits. Both Anthropic and Google now ship an agent that drives your own signed-in Chrome — Claude in Chrome and Chrome Auto Browse. If that is what you want, use them: they are free, first-party, and better integrated. They are also supervised — visible window, real time, and they hand control back to you at a login wall or a CAPTCHA. vibatchium is for the other half: unattended, headless, N-at-a-time, on a box with no human in front of it, against sites that fight automation. That is the whole of the wedge, and it is worth being precise about which side of it you are on.

pipx install vibatchium             # core: browse / extract / screenshot / N parallel sessions
# want the stealth HTTP fetch lane (vb fetch), the credential vault, VLM read, or the REST shim?
pipx install 'vibatchium[all]'      # everything; or pick extras: vibatchium[fetch], [secrets], [llm], [rest]
patchright install chrome
vb setup                    # register MCP + an auto-discoverable skill so agents reach for vb (idempotent)

Core install covers all browsing. vb fetch (the curl_cffi TLS-fingerprint lane) is the [fetch] extra; vb install reports which optional lanes are available. On a uv venv (no pip), add an extra with uv pip install --python <venv>/bin/python curl_cffi.

Bleeding edge from master: pipx install 'git+https://github.com/trueoriginlabs/vibatchium#egg=vibatchium[all]'

Coding agents (Codex / Cursor / Claude Code): read AGENTS.md first — it has the one-call recipes (explore, research) and the env-discovery traps to skip.

vb explore https://example.com                      # one-call: text-first (screenshot only as a fallback)
vb research --target https://example.com \          # parallel fan-out, N intents
  --intent "pricing model" --intent "customers" --intent "tech stack"

Status: active development, alpha. 1,061 tests green in CI (Linux, Python 3.11–3.13). Apache-2.0 (AGPL only via the opt-in nodriver extra).

_{Detector scores quoted below (bot.sannysoft, CreepJS, Cloudflare cold-launch) are manual observations, not CI-asserted — no test in the suite gates on them, and they are only as current as the last hand-run. The generated block under Measured scores is the one to trust; it is empty until someone runs it.}

Updating

vb update                  # upgrade to the latest PyPI release + restart the daemon
vb update --version 0.18.12  # or pin a specific version

vb update detects how vibatchium was installed (pipx, uv tool install, a pip-less uv venv, or pip with a PEP-668 --break-system-packages fallback) and then stops the running daemon so the next command loads the new code. Manual equivalent:

pipx upgrade vibatchium    # or: uv tool upgrade vibatchium / pip install -U vibatchium
vb shutdown                # bounce the daemon — it serves old code until you do
vb --version               # confirm

The daemon-restart step is the one people miss: the long-running daemon keeps serving the old version until it's bounced. vb update does it for you; if you upgrade by hand, run vb shutdown (the next vb call auto-respawns the new version). Optional features upgrade via pipx install 'vibatchium[all]' --force.

Why vibatchium

Most of what used to sit in this table is now commodity. What follows keeps only the rows that still separate us — and concedes the ones that don't.

_{Verified 2026-08-03 against agent-browser v0.33.2. — = not assessed, not a claim of absence.
¹ vercel-labs/agent-browser, ~1.27M npm downloads/week — by far the largest overlap, and it covers four of the rows above.
² Stealth is out of agent-browser core by policy; its docs designate a launch.mutate plugin capability as the extension point.
³ agent-browser's own README: "State files contain session tokens in plaintext … for encryption at rest, set AGENT_BROWSER_ENCRYPTION_KEY". Its credential vault is always encrypted; the session state is what defaults to plaintext.
⁴ agent-browser auto-generates its key into ~/.agent-browser/.encryption-key — the same directory as the ciphertext.
⁵ Scoped deliberately: this is a real differentiator against playwright-mcp, whose README notes a persistent profile can only be used by one browser instance at a time. It is not a differentiator against tools that give each named session its own browser instance.
⁶ Defaults to flag-only since 0.18.9: responses gain prompt_injection_risk + signals and content is left byte-identical, so nothing downstream changes. wrap / redact rewrite content and stay opt-in; VIBATCHIUM_SAFETY_MODE=off restores zero overhead.}

Real Chrome vs fake Chrome

A wave of "headless browser for AI agents" tools rebuild the browser from scratch (Rust + V8, no Blink/Skia) to hit tiny memory and sub-100ms page loads. The catch is structural: with no rendering engine, they can't produce a real device's fingerprint — they synthesize one. And synthetic fingerprints don't hold still.

vibatchium drives real Google Chrome, so its fingerprints are real — and, more to the point, stable. The single test that separates the two is fingerprint stability across navigations. Run the same canvas + WebGL probe on two pages in one session:

_{¹ Chrome's own software renderer (SwiftShader) by default — still a coherent, deterministic Chrome value, not a stub. A hardware-GPU string (e.g. ANGLE (Intel …)) needs the opt-in --gpu flag.}

A real device returns the same fingerprint every page load; a fingerprint keyed off Date.now() does not — and that inconsistency is exactly what lie-detection fingerprinters (CreepJS and friends) flag. Measured: vibatchium's canvas hash and WebGL readback are byte-identical across navigations, and CreepJS reported 0 % stealth-tampering (no synthetic-environment signatures) when last run by hand. That figure is not regression-tested — treat it as an observation, not a guarantee.

This is not a claim of invisibility. The moat is fingerprint authenticity, not hiding that a browser is automated — vibatchium still reads as headless on the headless-specific tells (see Honest limits), and real-GPU WebGL (--gpu) is opt-in. But real, consistent fingerprints pass the consistency tier that synthetic ones fail by construction — and that tier is what stands between you and a login wall.

Multi-session in 10 lines

vb session new work
vb --session work start
vb --session work go https://github.com           # log in by hand once
vb session new banking
vb --session banking start
vb --session banking go https://bank.example.com
vb --session work click @e3 &                     # truly parallel —
vb --session banking fill @e5 hi &                # separate Chromes, no cookie bleed
wait
vb session list

Active-session resolution: --session FLAG → $VIBATCHIUM_SESSION env → ~/.config/vibatchium/active-session → default. Cap via VIBATCHIUM_MAX_SESSIONS=8 (default 8).

Multi-agent: shared sessions vs a private daemon

On one shared daemon, sessions give real fingerprint isolation (separate Chromes, no cookie bleed) but share the host: the session count budget, the memory, and the blast radius of an OOM or a daemon bounce. Two models, pick per trust level:

• Cooperating agents (your own fleet): the shared daemon is right — just give each concurrent agent a unique --session name so stateful flows don't collide on default. vb session lease coordinates a shared name.
• A private blast radius: a per-agent daemon on its own socket + HOME — separate profiles/config/state, its own session budget, zero contact with the shared daemon. vb daemon start --isolated prints the XDG_RUNTIME_DIR/ HOME to export for subsequent calls; vb mcp --isolated runs the MCP server on its own private daemon directly. vb daemon reap cleans up abandoned ones. (Same UID = same trust domain — this bounds blast radius, not a security boundary between distrusting tenants; for that, separate UIDs/containers.)

Resource governance. The session cap bounds process count, not bytes. On a shared box, set VIBATCHIUM_SESSION_RAM_FLOOR_MB to refuse a new launch when free memory is low (a portable admission belt). For a hard ceiling, run the daemon under a cgroup — systemd-run --user --scope -p MemoryMax=4G vb daemon start puts the daemon and all its Chromes in one cgroup sharing the limit: an aggregate daemon-wide cap (not per-renderer), and a breach OOM-kills inside the scope, which can include the daemon. It's the only non-racy memory bound, so size it for the whole fan-out.

Idle CPU. Parked sessions can't burn cores either: the daemon SIGSTOPs a launched session's renderer processes after VIBATCHIUM_IDLE_FREEZE_AFTER seconds with no verb (default 90) and thaws them on the next call, so an idle WebGL / animation page drops to zero CPU without a teardown (default on; VIBATCHIUM_IDLE_FREEZE=0 disables).

Documentation

• AGENTS.md — coding-agent contract (Codex / Cursor / Claude Code)

Server modes

REST capability gating: vb serve --caps=core,nav,input,vision restricts the HTTP surface the same way mcp --caps does. Without it, REST grants local-code-equivalent access (eval + secret_* + file-writing verbs all exposed) — safe for localhost dev, not for hosted/multi-tenant.

Stealth tiers — what clears what

Stealth is a ladder, not a boolean. Pick the lowest tier that clears your target (higher tiers cost more setup / a visible browser / a manual login). vibatchium does not claim cold-launch defeat of behavioral walls — those need a real human-driven session, and attach-mode is the honest answer.

Architecture caveat. Every tier below was measured on x86-64 Linux. Patchright has a known open arm64 / Apple-Silicon detection gap, so these results should not be assumed to carry to an ARM host. If you run there, measure before you rely on it.

Measured scores

vb evals run --update-readme writes measured numbers into the block below, so what we publish is generated rather than asserted. It is empty until someone runs it — an empty block is honest; a number with no run behind it is not.

No eval run has been published yet. Generate with: vb evals run --update-readme

What these do and don't cover. These are fingerprint scoreboards — the static axis. Through 2026 the major anti-bot vendors moved to session-lifetime behavioural scoring, which none of these targets measure, and which we have not measured against any commercial vendor. Treat a good score here as evidence about environment coherence only.

For the behavioural axis itself, vb oracle run is a self-hosted probe: it drives a page with humanize off then on and grades trajectory curvature, dwell, keystroke cadence and scroll dynamics against a human-plausible band (vb oracle record captures a real-operator baseline; literature defaults until you do). It measures our model of human rather than a named vendor — but it turns "we humanize" into a measured on/off delta, and it's honest about the one axis synthetic input can't reach: CDP input emits no raw-pointer / coalesced events, which only attach-mode against real hardware closes.

Escalation ladder when a wall trips: headless → --headed → humanize on → --backend nodriver → attach-mode after a manual login. Patchright's CDP-layer patches apply in all tiers, including attach (connect_over_cdp).

The fetch verb is an orthogonal fast-path, not a tier: once you're past a wall in the browser, vb fetch reuses that session's cookies+proxy to hit JSON/API endpoints at TLS-fingerprint-correct speed — but it runs no JS, so it can't clear a JS challenge itself.

Attach mode — the practical Cloudflare workaround

For DataDome / Kasada / hardened auth that walls cold-launch automation:

google-chrome --remote-debugging-port=9222 \
              --disable-blink-features=AutomationControlled \
              --user-data-dir=/tmp/cdp-profile &
# log into the walled site by hand
vb attach http://localhost:9222
vb go https://target.example.com        # now reads as your real browser

Patchright's CDP-layer stealth still applies over connect_over_cdp — attach mode gets the same protocol-level patches as cold launch, plus your real-browser fingerprint and any cookies from the manual login.

Launch flags are yours on this tier. On cold launch the backend supplies --disable-blink-features=AutomationControlled for you. Attach connects to a Chrome that is already running, so nothing vibatchium does can add a launch flag after the fact — if you started Chrome without it, that tell is present for the whole session. Include it in the command above.

--remote-debugging-port is an open door. It grants full browser control to any process on the machine, and a page you visit can probe localhost to discover it. Use it on a machine you trust, and close Chrome when you're done.

Security model

vibatchium is built to drive real logins from an untrusted agent loop, so the threat model is "a credential must never reach the model, a screenshot, or a log":

• Encrypted vault. Passwords and TOTP secrets live in an XSalsa20-Poly1305 vault keyed from the OS keyring or VIBATCHIUM_SECRETS_KEY. A resolved secret never appears in logs, HAR captures, the observe cache, or any agent-visible response field (grep-tested in CI).
• Secrets are never rendered in the clear. fill --use-secret masks the field in the page (-webkit-text-security), applied before the value is written, so every path that turns the viewport into bytes — the screenshot verb, the 5 fps live-view stream, and VLM vision_* calls that ship the frame to a model — captures dots, not the value. The mask fails closed (no write if it can't be confirmed), covers password fields so a show-password toggle can't unmask, and the accessibility snapshot returned by map / diff_map strips masked values so the secret can't leak into the model's context as text either.
• Live-view is authenticated. The WebSocket requires a per-server token and rejects foreign-Origin connections (the CSWSH class), and driving the page is a separate token from watch-only — a read-only link can be shared without handing over the keyboard. Binds 127.0.0.1 by default (--insecure-public to override).
• Scraped content is marked untrusted. MCP verbs that return page-derived text carry openWorldHint, so a host can taint the output against prompt injection instead of treating a scraped page as instructions; pure probes are readOnlyHint and mutating verbs (stop, secret_delete, storage_restore) are destructiveHint.
• REST shim. Without --caps, the bearer token grants every verb including eval, secret_*, and file-writing verbs — local-code-equivalent, so always pass --caps=... in hosted mode. All vibatchium-written files are 0600; directories 0700.

Honest limits

• 5+ concurrent sessions = 1-2GB RAM. Each persistent-context Chrome is ~200-400MB. Bump cap with VIBATCHIUM_MAX_SESSIONS=8.
• Vision spend cap is process-wide. N fan-out agents share one daily/lifetime budget.
• Init scripts don't work on patchright backend. chrome.runtime stays undefined — accepted trade for stealth wins.
• Login walls (X, LinkedIn) require attach mode. Cold-launch fan-out can't defeat sites requiring authenticated sessions.
• Synthetic input has a CDP coordinate signature. Every click/type/hover/scroll rides Playwright over CDP Input.dispatchMouseEvent/dispatchKeyEvent (pageX==screenX, no CoalescedEvents). Patchright patches the JS-context leaks, not the Input domain, and humanize on improves trajectory/timing realism but does not change the per-event signature. Behavioral walls that fingerprint it (DataDome/Kasada/HUMAN) want attach-mode against a real headful Chrome you drive — OS-level synthetic input (CDP-Patches) is headful + active-tab only and doesn't fit a headless, N-parallel daemon.
• fetch is a static-fingerprint lane, not a browser. The curl_cffi fetch verb matches Chrome's JA3/HTTP2 but runs no JavaScript — it clears TLS-fingerprint gates, not DataDome/Kasada/Turnstile JS challenges. Fall back to go for those.
• Single daemon = single point of failure. No HA built in.
• Behavioural detection now targets the humanizer directly. Cloudflare's 2026 bot-detection work names mathematically ideal Bézier cursor paths and superhuman click precision as tells. humanize improves on nothing-at-all, but it is a pointer-trajectory model, not a physiological one — and it is off by default, which on a behaviourally-scored site is the louder of the two states.
• One burned profile can taint every account that shares it. Vendors now link device telemetry across sessions and accounts. Use one profile per account, never share a profile between identities, and don't reuse a profile that has already been challenged.

Authorized use

vibatchium is built to drive sessions you own, with your credentials, on your machine — your accounts, your employer's, or a client's with their written permission. It is a tool for automating access you already have.

That boundary is not a formality. In 2026 a US district court granted a preliminary injunction against an AI agent that accessed password-protected pages through the user's own logged-in account, holding that the user's permission is not the platform's authorization. (The order was stayed on appeal and there is no merits ruling, so the law here is unsettled — which is a reason for care, not comfort.) Scraping a site's public pages, evading a wall you have no account behind, or automating an account whose terms forbid it are all decisions you are making, and the consequences are yours.

Check the terms of the site you are automating. If you are acting for someone else, get it in writing.

License

Apache-2.0 core. Every default-install extra is permissive too — the fetch lane's curl_cffi is MIT. The only copyleft option is the opt-in nodriver backend (AGPL-3.0) — consult licensing before integrating it commercially. Nothing GPL/AGPL ships in the base install or [all].