VulpineOS
Star on GitHub →
Open source · MPL 2.0

The runtime thatmakes browseragents finally work.

VulpineOS runs AI agents that navigate the web as reliably as a human, blending in, handling surprises, and cutting through the noise. Open source. Self-hosted. Drop in your LLM keys and go.

Read the docs →Star on GitHub
The problem

Browsers were not designed for AI.
Your agents pay the price.

PROBLEM 01PROMPT INJECTION

Your agent follows instructions it should never see.

Websites hide text in elements the human never notices. Your agent reads everything and attackers exploit that to rewrite its instructions mid-task.

→ one hidden paragraph can hijack your agent
PROBLEM 02PAGE MUTATION

The page your agent sees is already stale.

Between reading the page and taking action, timers fire, elements shift, modals appear. Your agent acts on a page that no longer exists.

→ your agent acts on a page that no longer exists
PROBLEM 03TOKEN WASTE

Most of what your agent reads is useless.

Accessibility trees are full of scaffolding and wrappers. Your model pays per token to read a page you already understand. The cost adds up fast.

→ you pay for reasoning, not page scaffolding
Architecture

Four layers of reliability. One pipeline from page to prompt.

Built at the browser level so your agent never has to work around browser limitations.

INPUThostile web page
OUTPUTsafe, compact snapshot
PHASE 01 / 04
INJECTION FILTER

Strip everything the agent should not see.

Hidden text, invisible elements, offscreen content. Anything the human would not encounter gets removed before the agent ever reads it.

hidden textinvisible elementsoffscreen content
Stealth & network

Your agent looks
like a person on every site.

Every session the site sees a unique browser profile. Consistent across WebGL, fonts, timezone, locale. Bring your own proxies and the runtime keeps everything in sync automatically.

Unique identity per session
WebGL, canvas, audio, fonts, and timezone stay consistent across the stack
Your proxies, auto-synced
Plug in your own residential pool. Geolocation, timezone, locale, and WebRTC IP all re-sync on each rotation.
Pre-warmed profiles
Profiles age across innocuous sites before hitting the target
Network-level cover
JA3/JA4 spoofed to match the fingerprinted browser identity
Fleet at scale

Run a fleet,
not a script.

A fleet of agents working in parallel, each with its own identity, instructions, and budget. Manage them from one panel or let cost alerts and webhooks handle the monitoring.

CONCURRENT
~200
Active agents on a 32GB / 16-core workstation. Scales with your hardware: more on bigger nodes, less on a laptop. Bring your own LLM and proxies.
ORCHESTRATION
A2A bus
Agent-to-agent messaging built in. One agent can spawn, supervise, or hand off to another, with cost ceilings and approval gates.
LLM PROVIDERS
30+
Anthropic, OpenAI, Google, xAI, Groq, Mistral, Cerebras, OpenRouter, Ollama, and many more. Drop in your API key and switch providers per agent.
MCP TOOLS
36
Browser navigation, clicks, downloads, screenshots, audio, and mobile are exposed as MCP tools that any LLM can call directly.
Token savings

Pay for reasoning,
not scaffolding.

Most browser snapshots are full of wrappers, separators, and repetition. VulpineOS strips everything the model does not need. You keep the structure and the text, not the scaffolding.

Chrome full AX tree~42,832 tokens
Raw page HTML~12,761 tokens
Playwright ariaSnapshot~11,577 tokens
VulpineOS optimized export~2,942 tokens
93.1%
Measured reduction vs compact Chrome AX tree
74.6%
Measured reduction vs Playwright ariaSnapshot
39.9M
Input tokens avoided per 1,000 snapshots vs compact Chrome AX
The landscape

Everything your
agents need. Nothing they don’t.

Other tools give you a browser and call it a day. VulpineOS gives you a browser runtime tuned for AI agents, with protections that cannot be bolted on later.

FeatureVulpineOSPlaywrightPuppeteerBrowserbaseBrowserless
C++ injection-proof a11y filter————
Action-Lock page freeze (nsDocShell)————
Token-optimized DOM export——PARTIAL—
Anti-detect fingerprints———
Built-in proxy rotation (BYO pool)——PARTIAL—
Self-host runtime——
Multi-agent orchestration built in————
Open source license——
Comparisons as of April 2026 · based on publicly-available product documentation.
Features

What you can
do with it.

01
Invisible to anti-bot systems
Every session looks like a real person on a real device. Consistent fingerprints across WebGL, fonts, and timezone.
02
Your agent follows your instructions
Hidden text and injected prompts are stripped before the agent ever sees them.
03
Reliable actions, every time
The page freezes while your agent decides. No timing issues, no surprise modals.
04
Smaller prompts, lower costs
Snapshots are 90% smaller than standard browser output. You pay for reasoning, not markup.
05
Structured data, not guesswork
Your agent returns typed, validated JSON. No parsing, no surprises.
06
Bring your own proxies
Use the pool you already have. The runtime keeps location and fingerprints in sync through every rotation.
07
Any model you want
Pick the best LLM for each task. Switch providers per agent without changing your setup.
08
36 tools out of the box
Navigation, clicks, downloads, screenshots, audio, mobile. Every action is an MCP tool your agent can call.
09
Control your spend
Per-agent budgets with soft and hard limits. Webhooks when costs cross your thresholds.
Built in the open

Built in public.
Free to use.

The whole stack is open source under MPL 2.0. Fork it, extend it, build on it. Contribute upstream or keep your changes private.

FOXBRIDGE

Plug Firefox-based browsers into anything that speaks CDP.

An open-source protocol translator exposes a Chrome DevTools Protocol surface on top of Juggler and WebDriver BiDi. OpenClaw and any CDP-native agent stack drives Camoufox unchanged, with full anti-fingerprint isolation per session.

$ foxbridge start --browser camoufox --port 9222

  ✓ Camoufox 146.0.1 detected
  ✓ Juggler bridge online
  ✓ WebDriver BiDi backend ready

  CDP endpoint
  ws://localhost:9222/devtools/browser

  drive with: openclaw, playwright, puppeteer, …
Clone, build, ship

Ship your first
agent today.

Clone the repo, add your LLM keys, and run. All under MPL 2.0.

Read the docsStar on GitHub
Open source underMPL 2.0Forked from CamoufoxBuilt on Firefox 146