VFA — Visual & Functional Web Auditor (MCP Server)

Autonomous black-box web QA agent exposed as an MCP server over stdio. Pure hexagonal architecture (Ports & Adapters), one unified inference engine powered by LiteLLM, and a local-first Playwright Chromium sandbox — no Docker, no sidecars, no infrastructure to maintain.

Why VFA

Pillar	What it means
Native Zero-Docker	Playwright downloads and manages Chromium on the host itself. A built-in ephemeral BrowserContext sandbox plus an Anti-SSRF firewall ship by default, and local development is first-class (localhost, 127.0.0.1, *.local).
Universal LiteLLM Engine	A single adapter drives DeepSeek, OpenAI, Anthropic, Gemini, xAI/Grok, OpenRouter and Ollama. Switching providers is a two-variable change.
Active Visual Grounding	Multimodal vision inspects real UIs, reads raster image CAPTCHAs and clicks by visual coordinates whenever the accessibility tree is not enough.
1-Click Execution	Plug-and-play: uvx vfa-mcp starts the server with no permanent install, pip install vfa-mcp if you prefer an isolated virtualenv.

uvx vfa-mcp

Quickstart

⚠️ License Key (Required)

VFA is a commercial tool. An active VFA Pro License is required to execute audits and user flows.

• Get your license: Purchase on Creem
• Configuration: Set the VFA_LICENSE_KEY environment variable in your MCP client or .env.

1. Install

One command, permanent install in your own virtualenv:

pip install vfa-mcp && playwright install chromium

Zero-install alternative (recommended for MCP clients):

uvx vfa-mcp
playwright install chromium   # first run only

2. Pick a provider

VFA resolves the API key from VFA_LLM_API_KEY first, then from the standard variable of the selected provider (OPENAI_API_KEY, ANTHROPIC_API_KEY, GEMINI_API_KEY, DEEPSEEK_API_KEY, XAI_API_KEY, OPENROUTER_API_KEY).

Provider	VFA_LLM_PROVIDER	Example VFA_LLM_MODEL	API key variable	Base URL / Endpoint variable
OpenAI	openai	gpt-4o	OPENAI_API_KEY	— (Default)
Anthropic	anthropic	claude-3-5-sonnet-20241022	ANTHROPIC_API_KEY	— (Default)
Google Gemini	gemini	gemini/gemini-2.0-flash	GEMINI_API_KEY	— (Default)
DeepSeek (Direct)	deepseek	deepseek/deepseek-chat	DEEPSEEK_API_KEY	— (Default)
DeepSeek Vision	deepseek	deepseek/deepseek-flash	DEEPSEEK_API_KEY	— (Default)
DeepSeek (OpenAI-compat)	openai	openai/deepseek-chat (o deepseek-flash)	DEEPSEEK_API_KEY / OPENAI_API_KEY	OPENAI_API_BASE="https://api.deepseek.com"
DeepSeek (Anthropic-compat)	anthropic	claude-3-5-sonnet (alias)	DEEPSEEK_API_KEY / ANTHROPIC_API_KEY	ANTHROPIC_BASE_URL="https://api.deepseek.com/anthropic"
xAI / Grok	xai	xai/grok-2-vision-1212	XAI_API_KEY	— (Default)
OpenRouter	openrouter	openrouter/anthropic/claude-3.5-sonnet	OPENROUTER_API_KEY	— (Default)
Ollama (local)	ollama	ollama/llama3.2-vision	— (Placeholder)	OLLAMA_API_BASE="http://localhost:11434"

Aliases accepted by the provider resolver: google (Gemini) and grok (xAI). Vision falls back to the reasoning model when VFA_VISION_PROVIDER / VFA_VISION_MODEL are left empty.

3. Wire it into your client

Zoo Code / Roo Code (mcpSettings.json)

Click the MCP Servers plug icon ⚙️ → Edit MCP Settings, then add:

{
  "mcpServers": {
    "vfa": {
      "command": "uvx",
      "args": ["vfa-mcp"],
      "env": {
        "VFA_LICENSE_KEY": "PRO-XXXX-XXXX",
        "OPENAI_API_KEY": "sk-...",
        "HEADLESS": "true",
        "VFA_ALLOW_LOCAL_HOST": "true",
        "LITELLM_LOG": "ERROR"
      }
    }
  }
}

---

Claude Desktop (claude_desktop_config.json)

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "vfa": {
      "command": "uvx",
      "args": ["vfa-mcp"],
      "env": {
        "OPENAI_API_KEY": "sk-...",
        "HEADLESS": "true",
        "VFA_ALLOW_LOCAL_HOST": "true",
        "LITELLM_LOG": "ERROR"
      }
    }
  }
}

---

Cursor (.cursor/mcp.json or Settings → Features → MCP)

{
  "mcpServers": {
    "vfa": {
      "command": "uvx",
      "args": ["vfa-mcp"],
      "env": {
        "OPENAI_API_KEY": "sk-...",
        "HEADLESS": "true",
        "VFA_ALLOW_LOCAL_HOST": "true",
        "LITELLM_LOG": "ERROR"
      }
    }
  }
}

---

Provider Drop-in Presets (Replace the "env" block)

Simply swap the "env" object in any configuration above with your preferred model:

⚡ DeepSeek (Reasoning + Fast Vision)

"env": {
  "VFA_LLM_PROVIDER": "deepseek",
  "VFA_LLM_MODEL": "deepseek-chat",
  "VFA_VISION_PROVIDER": "deepseek",
  "VFA_VISION_MODEL": "deepseek-flash",
  "DEEPSEEK_API_KEY": "sk-...",
  "HEADLESS": "true",
  "VFA_ALLOW_LOCAL_HOST": "true",
  "LITELLM_LOG": "ERROR"
}

🧠 Anthropic Claude 3.5 Sonnet

"env": {
  "VFA_LLM_PROVIDER": "anthropic",
  "VFA_LLM_MODEL": "claude-3-5-sonnet-20241022",
  "ANTHROPIC_API_KEY": "sk-ant-...",
  "HEADLESS": "true",
  "VFA_ALLOW_LOCAL_HOST": "true",
  "LITELLM_LOG": "ERROR"
}

🚀 Google Gemini 2.0 Flash (Multimodal)

"env": {
  "VFA_LLM_PROVIDER": "gemini",
  "VFA_LLM_MODEL": "gemini/gemini-2.0-flash",
  "GEMINI_API_KEY": "AIzaSy...",
  "HEADLESS": "true",
  "VFA_ALLOW_LOCAL_HOST": "true",
  "LITELLM_LOG": "ERROR"
}

🌐 OpenRouter (Universal Gateway)

"env": {
  "VFA_LLM_PROVIDER": "openrouter",
  "VFA_LLM_MODEL": "openrouter/anthropic/claude-3.5-sonnet",
  "VFA_VISION_MODEL": "openrouter/google/gemini-2.0-flash-001",
  "OPENROUTER_API_KEY": "sk-or-v1-...",
  "HEADLESS": "true",
  "VFA_ALLOW_LOCAL_HOST": "true",
  "LITELLM_LOG": "ERROR"
}

🏠 Ollama (100% Local / Offline)

"env": {
  "VFA_LLM_PROVIDER": "ollama",
  "VFA_LLM_MODEL": "ollama/llama3.2-vision",
  "OLLAMA_API_BASE": "http://localhost:11434",
  "HEADLESS": "true",
  "VFA_ALLOW_LOCAL_HOST": "true",
  "LITELLM_LOG": "ERROR"
}

⚙️ OpenAI-compatible:

"env": {
  "VFA_LLM_PROVIDER": "openai",
  "VFA_LLM_MODEL": "openai/deepseek-v4-pro",
  "VFA_VISION_PROVIDER": "openai",
  "VFA_VISION_MODEL": "openai/deepseek-flash",
  "OPENAI_API_BASE": "https://api.deepseek.com",
  "OPENAI_API_KEY": "sk-...",
  "HEADLESS": "true",
  "VFA_ALLOW_LOCAL_HOST": "true",
  "LITELLM_LOG": "ERROR"
}

⚙️ Anthropic-compatible:

"env": {
  "VFA_LLM_PROVIDER": "anthropic",
  "VFA_LLM_MODEL": "claude-3-5-sonnet",
  "ANTHROPIC_BASE_URL": "https://api.deepseek.com/anthropic",
  "ANTHROPIC_API_KEY": "sk-...",
  "HEADLESS": "true",
  "VFA_ALLOW_LOCAL_HOST": "true",
  "LITELLM_LOG": "ERROR"
}

Agent-facing usage guide (when to invoke each tool, report template, security rules): SKILL.md. Marketplace manifest: skill.json.

Tool reference

qa_audit_url(url)

Full-page screenshot, vision analysis and runtime error collection in a single call.

Parameter	Type	Required	Description
url	string	yes	Reachable URL to audit. Private/loopback targets are rejected unless VFA_ALLOW_LOCAL_HOST is enabled.

Returns url, ok (true only when the audit passed and no runtime event was captured), summary, screenshot_base64 (PNG), issues[] (visual, accessibility and layout findings) and errors[].

qa_execute_user_flow(url, steps)

Replays a user journey described in plain language (English or Spanish).

Parameter	Type	Required	Description
url	string	yes	Starting URL of the flow.
steps	string[]	yes	Ordered natural-language steps, e.g. ["Click 'Sign in'", "Verify ..."].

Returns url, success, ok, steps[{step, ok, detail}], screenshots[] (base64 PNG milestones) and errors[].

qa_get_runtime_errors()

Drains the browser error buffer: console.error, uncaught exceptions (pageerror) and HTTP responses with status >= 400.

Parameter	Type	Required	Description
—	—	—	No arguments.

Returns count and errors[{kind, message, url, status}], where kind is one of console, pageerror or http. The call is read-destructive: it empties the buffer, so invoke it once at the end of a sequence.

Configuration reference

VFA is configured exclusively through environment variables — there are no config files, no database and no flags to pass. Variables are grouped in three independent layers:

Layer	Variables	Consumed by
Reasoning engine	VFA_LLM_PROVIDER, VFA_LLM_MODEL, VFA_LLM_API_KEY, VFA_LLM_BASE_URL	The Deep Agent (LiteLLM tool-calling loop)
Vision engine	VFA_VISION_PROVIDER, VFA_VISION_MODEL, VFA_VISION_API_KEY	Visual grounding / CAPTCHA fallback
Browser & runtime	HEADLESS, VFA_ALLOW_LOCAL_HOST, BROWSER_CDP_*, VFA_SESSION_*, …	Playwright adapter, session pool, Anti-SSRF

Is anything mandatory?

Only two things are strictly mandatory:

1. VFA_LICENSE_KEY: Your commercial license key purchased from Creem.
2. An inference API key for your chosen provider (e.g. OPENAI_API_KEY, DEEPSEEK_API_KEY, etc.).

# Minimal production setup — works because the defaults are openai/gpt-4o
export OPENAI_API_KEY="sk-..."

Does VFA_LLM_MODEL drive the Deep Agent?

Yes. VFA_LLM_MODEL is the model used by the Deep Agent — the component that executes qa_execute_user_flow and performs the reasoning step of qa_audit_url. Both tools go through LiteLLM with VFA_LLM_PROVIDER + VFA_LLM_MODEL, so:

• You may omit them when you use OpenAI gpt-4o and export OPENAI_API_KEY (defaults: openai / gpt-4o).
• You must set them when you use any other provider, any other model, or an OpenAI-compatible gateway (DeepSeek, OpenRouter, Ollama, vLLM, Azure proxy…).
• The model must support tool calling (native function calling), because the ReAct loop sends a JSON action schema on every iteration.

For OpenAI-compatible endpoints, pair the model with a base URL:

export VFA_LLM_PROVIDER="openai"          # protocol, not the vendor
export VFA_LLM_MODEL="openai/deepseek-chat"
export OPENAI_API_KEY="sk-..."            # or VFA_LLM_API_KEY
export OPENAI_BASE_URL="https://api.deepseek.com"

Credential and endpoint precedence

What	Resolution order
Reasoning API key	VFA_LLM_API_KEY → provider standard variable (OPENAI_API_KEY, ANTHROPIC_API_KEY, GEMINI_API_KEY, DEEPSEEK_API_KEY, XAI_API_KEY, OPENROUTER_API_KEY) → OPENAI_API_KEY
Reasoning base URL	VFA_LLM_BASE_URL → OPENAI_BASE_URL → OLLAMA_API_BASE (only when the provider is ollama)
Vision provider	VFA_VISION_PROVIDER → VFA_LLM_PROVIDER → openai
Vision model	VFA_VISION_MODEL → VFA_LLM_MODEL → gpt-4o
Vision API key	VFA_VISION_API_KEY → VFA_LLM_API_KEY → provider standard variable → OPENAI_API_KEY

VFA_VISION_* is therefore optional by design: if the reasoning model is multimodal (gpt-4o, claude-3-5-sonnet, gemini/gemini-2.0-flash, xai/grok-2-vision-1212), leave it empty and one model does everything. Set it only when you want a cheaper/faster model for screenshots or when your reasoning model is text-only.

Full variable list

Variable	Default	Required when	Purpose
VFA_LICENSE_KEY	(empty)	Mandatory for all tool executions	Commercial Pro license key from Creem
VFA_LLM_PROVIDER	openai	Using a non-OpenAI provider	SDK protocol/routing for the Deep Agent (openai, anthropic, gemini, deepseek, xai, openrouter, ollama; aliases google, grok)
VFA_LLM_MODEL	gpt-4o	Using a model other than gpt-4o	Reasoning model (must support tool calling)
VFA_LLM_API_KEY	(empty)	Sharing one credential across providers	Generic key that overrides the provider standard variable
VFA_LLM_BASE_URL	(empty)	OpenAI-compatible gateways, self-hosted or proxied endpoints	Base URL override (OPENAI_BASE_URL also honoured)
VFA_VISION_PROVIDER	inherits	Vision served by a different vendor	Routing for multimodal calls
VFA_VISION_MODEL	inherits	Reasoning model lacks image input, or cost split desired	Model used for screenshots and visual grounding
VFA_VISION_API_KEY	inherits	Vision served by a different vendor	Credential for multimodal calls
OLLAMA_API_BASE	(empty)	VFA_LLM_PROVIDER=ollama	Local endpoint, e.g. http://localhost:11434 (no key needed, a placeholder is used)
VFA_AGENT_MAX_ITERATIONS	12	Long flows that exceed the ReAct budget	Iteration ceiling of the Deep Agent loop
VFA_LLM_TEMPERATURE	0.0	Never (deterministic QA by default)	Sampling temperature
VFA_LLM_TIMEOUT	60	Slow/self-hosted models	Per-call timeout in seconds
VFA_LLM_REQUESTS_PER_SECOND	0	Free tiers or strict quotas	Token-bucket rate limiter; 0 disables it
VFA_LLM_CHECKS_PER_SECOND	10.0	Never	Rate limiter polling frequency
HEADLESS	true	Local debugging of the flow step by step	Run Chromium without a window
VFA_ALLOW_LOCAL_HOST	true	Must be false for internet-facing production deployments	Allow localhost, 127.0.0.1, 0.0.0.0, *.local; false restores strict Anti-SSRF blocking of loopback, private ranges and cloud metadata
BROWSER_CDP_URL	(empty)	Attaching to an already running Chromium	Optional custom CDP endpoint; empty uses the managed local browser
BROWSER_CDP_TIMEOUT_MS	2000	Remote CDP endpoint slower than 2s	Timeout of the CDP connection attempt
VFA_ACTION_TIMEOUT_MS	10000	Heavy SPAs with slow first paint	Timeout of every individual Playwright action
VFA_SCREENSHOT_FULL_PAGE	true	Never (viewport-only reports needed)	Full-page screenshots instead of viewport only
VFA_SESSION_POOL_ENABLED	true	Stateless/multi-tenant deployments may disable it	Reuse browser sessions between tool calls
VFA_SESSION_POOL_TTL	300	Never	Idle seconds before a pooled session is closed
VFA_SESSION_POOL_MAX_SIZE	5	High-concurrency hosts	Maximum concurrent pooled sessions (LRU eviction)
VFA_SESSION_STORAGE_STATE	(empty)	Replaying an authenticated journey	Imported Playwright storageState; empty keeps cookies in memory only
SSL_VERIFY	true	Self-signed certificates in local/staging	TLS verification switch (leave empty to skip verification; any non-empty value keeps it on)

A copy-ready template with the same defaults lives in .env.example.

Production checklist

1. One credential per deployment, injected by your secret manager — never committed and never placed in the MCP client JSON for shared machines.
2. Keep VFA_ALLOW_LOCAL_HOST=false on any host that can reach an internal network; it is the only guard against SSRF from an LLM-driven navigation.
3. Pin VFA_LLM_MODEL explicitly instead of relying on gpt-4o; upstream providers rotate model names and a silent change alters QA verdicts.
4. Set VFA_LLM_REQUESTS_PER_SECOND when you run parallel audits against a rate-limited provider, and raise VFA_LLM_TIMEOUT for slow self-hosted models.
5. Keep HEADLESS=true on servers (no X server is available) and keep SSL_VERIFY enabled outside staging.
6. Do not mount VFA_SESSION_STORAGE_STATE unless the flow needs a reused login: it writes real session cookies to the host filesystem.

Architecture

app/
  domain/                  # Pure domain (no I/O, no frameworks)
    models.py              # AuditReport, StepResult, RuntimeErrorEvent, VFAContext
    parser.py              # Deterministic NLP parser (ES/EN)
    semantic.py            # Cascade: regex -> a11y snapshot -> LLM fallback
    vision.py              # Visual primitives and grounding contracts
  ports/                   # Abstract contracts
    browser.py             # BrowserSessionPort / BrowserGatewayPort
    llm.py                 # LLMPort / UsageLimiterPort
  adapters/
    browser/
      playwright_adapter.py  # Local sandboxed Chromium + ephemeral context
      network_guard.py       # Anti-SSRF firewall (private ranges / loopback)
      session_pool.py        # TTL + LRU pool, cookies in memory only
      capture.py             # console.error, pageerror, HTTP >= 400
      visual_primitives.py   # Screenshot and coordinate helpers
    llm/
      litellm_adapter.py     # Multi-provider acompletion + base64 vision
      deep_agent.py          # ReAct loop with native tool calling
      agent_tools.py         # Agent tool JSON schemas
      vision_grounding.py    # Visual grounding adapter
      rate_limiter.py        # Async token bucket
      visual_fallback.py     # Heuristic fallback chain
  tools/qa_tools.py        # The three QA tools
  config.py                # Single source of truth for configuration
server_mcp.py              # FastMCP entrypoint (stdio)

Security model

• Chromium starts with strict isolation flags (--disable-background-networking, --disable-default-apps, --disable-sync, --no-first-run, --hide-scrollbars).
• Every session uses an ephemeral in-memory BrowserContext, destroyed on close: no cache, cookies or storage are persisted to the host disk.
• A context-level Anti-SSRF firewall aborts (accessdenied) requests to loopback, private ranges (10/8, 172.16/12, 192.168/16) and cloud metadata (169.254.169.254). Local development targets are allowed only while VFA_ALLOW_LOCAL_HOST is true; set it to false for strict production mode.
• The LLM credential never reaches the audited page: only the server process talks to the inference provider.

Development

pip install -e ".[dev]"
playwright install chromium
pytest

Tests run without a real browser or network: every adapter is replaced by an in-memory double (tests/conftest.py, tests/visual_fakes.py).

License

Commercial, single-seat Pro license. Distributed on PyPI as vfa-mcp.
Acquire your permanent license key at Creem Checkout.