What is Visus and its primary purpose?

Visus is a pre-filter tool designed to sanitize web content for Large Language Models (LLMs). It removes noise, boilerplate, and ads, redacts PII, and detects prompt injection attempts before content reaches your LLM.

How does Visus help reduce LLM token consumption and costs?

By stripping irrelevant content like navigation, ads, and tracking scripts from web pages, Visus can reduce token consumption for LLMs by up to 70%, leading to significant cost savings.

What kind of security threats does Visus protect against?

Visus offers advanced protection against 43 categories of prompt injection attacks, including role hijacking, data exfiltration, encoding obfuscation, and direct instruction injection, generating compliance-grade threat reports.

Does Visus handle Personally Identifiable Information (PII)?

Yes, Visus automatically redacts common PII types such as email addresses, phone numbers, SSNs, credit card numbers, and IP addresses, ensuring sensitive data doesn't enter the LLM's context window or conversation history.

How does Visus integrate with AI agents like Claude?

Visus integrates as a local pre-filter, typically via `npx visus-mcp`, providing its robust web content sanitization, PII redaction, and prompt injection detection tools directly to your LLM agent for secure web access.

Visus: Secure LLM Web Access, Filter Noise & Detect Injection

Name: Visus
Author: visus-mcp

Visus is a crucial pre-filter designed to enhance the security and efficiency of AI agents interacting with web content. It strips away extraneous elements like navigation bars, ads, cookie banners, and tracking scripts, significantly reducing token consumption on bloated pages by up to 70%. Beyond noise reduction, Visus actively detects and neutralizes a wide array of prompt injection attempts and automatically redacts sensitive Personal Identifiable Information (PII) before it reaches the LLM's context window. This infrastructure-level tool ensures agents only process clean, relevant data, complemented by compliance-grade audit logs and human-in-the-loop security for critical threats.

主な機能

01Comprehensive Web Content Sanitization: Strips noise, ads, and boilerplate from web pages, reducing token consumption for LLMs by up to 70%.

02Advanced Prompt Injection Detection: Identifies and neutralizes 43 categories of injection attempts, including role hijacking, data exfiltration, and encoding obfuscation.

03Automated PII Redaction: Automatically redacts emails, phone numbers, SSNs, credit card numbers, and IP addresses before content reaches the LLM.

04Compliance-Grade Threat Reporting: Generates detailed reports aligned with OWASP LLM Top 10, NIST AI 600-1, MITRE ATLAS, and ISO/IEC 42001 standards.

05Multiple Web Content Extraction Modes: Offers `visus_fetch` for full page sanitization, `visus_read` for article content extraction, `visus_search` for sanitized DuckDuckGo results, and `visus_fetch_structured` for schema-based data extraction.

060 GitHub stars

ユースケース

01Protecting LLMs from malicious web content and prompt injection attacks.

02Optimizing LLM token usage by delivering clean, relevant web page content.

03Ensuring PII privacy and generating audit logs for AI interactions with external data sources.