Sunglasses MCP¶

A local input firewall that screens everything your agent reads. Sunglasses runs offline on the operator's machine and exposes MCP tools that scan text and files for prompt injection, credential leaks and data exfiltration against 1,554 patterns across 118 categories. A third tool reports the active pattern set, so operators know exactly what the scanner is enforcing.

Server type: Local (stdio, self-hosted)
Auth: none (local process)
Endpoint: python -m sunglasses.mcp
Tools: 3 (scan_text, scan_file, scanner_info)
Pricing: free and open source
Category: Security
Built by: Sunglasses (github.com/sunglasses-dev/sunglasses)

Why This Matters for Operators¶

Agents read untrusted content constantly: web pages, documents, tickets, emails. Prompt injection hides in that content, and a compromised agent with access to business tools is a real loss vector. Sunglasses gives operators a local, offline gate that screens text and files before the agent acts on them, without sending any data to a third-party scanner.

For operators who handle client or regulated data, the offline property is the point: the scan runs entirely on the machine, so documents never leave the environment to be checked for leaks.

Tools & Capabilities¶

Tool	Purpose
scan_text	Screen a text block for injection, leaks and exfiltration patterns
scan_file	Screen a file's contents with the same pattern set
scanner_info	Report the active pattern set and version

Installation¶

pip install sunglasses
python -m sunglasses.mcp

Start the module and connect it to any MCP client as a stdio server.

Configuration¶

{
  "mcpServers": {
    "sunglasses": {
      "command": "python",
      "args": ["-m", "sunglasses.mcp"]
    }
  }
}

Business Relevance¶

Ops teams gate agent input against injection and leak patterns
Security leads get a local, offline screening layer with no data egress
Agencies handling client documents scan before the agent processes files
Founders add defense in depth around agents with business tool access

Integration with CorpusIQ¶

Sunglasses fits the CorpusIQ security posture: before external content from email connectors or web fetches enters an agent workflow, scan_text can screen it. For CorpusIQ-powered agent deployments, Sunglasses is a complementary local layer to the guardrails doctrine, screening untrusted input before any external action fires.

Limitations¶

Pattern-based: novel injection techniques may slip past static rules
Local stdio server; each agent host needs its own installation
Single-purpose security tool, not a full agent security platform
New listing with no long track record

FAQ¶

Does any data leave the machine?¶

No. Sunglasses runs fully offline as a local process.

What does it scan for?¶

Prompt injection, credential leaks and data exfiltration across 1,554 patterns in 118 categories.

What are the three tools?¶

scan_text, scan_file and scanner_info, the last reporting the active pattern set.