Skip to main content
Glama
MCP Content Credentials Server

MCP Content Credentials Server

by noga7

MCP Content Credentials Server

MCP (Model Context Protocol) server for reading C2PA Content Credentials from images and videos. Detects credentials from both embedded manifests and invisible watermarks.

Features

  • πŸ” Embedded C2PA Detection - Read manifests from file metadata

  • 🌊 TrustMark Watermark Detection - Detect credentials in image pixels (survives social media!)

  • 🌐 URL Support - Check credentials from web URLs

  • πŸ“‚ Direct Filesystem Access - Claude can browse your directories

  • ⚑ Smart Detection - Checks embedded first, watermark as fallback

  • πŸ€– Automatic Installation - Zero configuration setup

  • πŸ“‹ Structured Output - Human-readable parsed data

  • πŸ›‘οΈ Production Ready - Full error handling and logging

  • 🌐 REST API - HTTP endpoints for ChatGPT and web integration

Related MCP server: chronoverify-mcp

Quick Start

# 1. Clone
git clone https://github.com/noga7/mcp-content-credentials.git
cd mcp-content-credentials

# 2. Install (automatic: installs c2patool + TrustMark)
npm install

# 3. Build
npm run build

# 4. Configure Claude Desktop
# Add to ~/Library/Application Support/Claude/claude_desktop_config.json:
{
  "mcpServers": {
    "content-credentials": {
      "command": "node",
      "args": ["/absolute/path/to/mcp-content-credentials/build/index.js"]
    }
  }
}

# 5. Restart Claude Desktop

# Optional: Start REST API for ChatGPT/web access
npm run start:api
# Server runs on http://localhost:3000

REST API (for ChatGPT & Web Apps)

Want to use this with ChatGPT or your own web app? Start the HTTP REST API:

npm run start:api

The server runs on http://localhost:3000. See REST-API.md for full documentation.

For ChatGPT: Use ngrok to expose your local server, or deploy to Render/Railway. See REST-API.md for instructions.

# Quick test
curl "http://localhost:3000/verify-url?url=https://example.com/image.jpg"

Prerequisites

  • Node.js v18+

  • Python 3.8.5+ (for TrustMark watermarks)

All other dependencies auto-install during npm install:

  • βœ… c2patool (Homebrew on macOS, binary on Linux)

  • βœ… TrustMark Python package (via pip)

Manual Installation (if auto-install fails)

# c2patool
brew install contentauth/tools/c2patool  # macOS

# TrustMark
pip3 install trustmark Pillow

# Or retry auto-install
npm run install-deps

Usage

Check a Specific File

"Check content credentials in ~/Desktop/photo.jpg"
"Is this image AI-generated?"
"Who created /Users/you/Downloads/image.png?"

Browse Directories

"What images are in my Desktop?"
"Check my Downloads for Content Credentials"
"Find AI-generated images in my Pictures"

Check URLs

"Check credentials at https://example.com/image.jpg"

How It Works

Detection Flow

1. Check Embedded C2PA Manifest (fast: ~150ms)
   ↓
   Found? β†’ Return immediately βœ…
   ↓
2. Check TrustMark Watermark (slower: ~600ms)
   ↓
   Found? β†’ Return watermark data βœ…
   ↓
3. Neither found β†’ "No Content Credentials found" ❌

Why This Order?

  • Performance: 80% of credentialed images have embedded manifests

  • Speed: Skip expensive watermark check when not needed

  • Completeness: Still catch stripped metadata via watermarks

TrustMark Watermarks

Invisible watermarks embedded in image pixels that:

  • βœ… Survive JPEG compression

  • βœ… Persist through social media uploads (Instagram, Twitter)

  • βœ… Work after print-scan cycles

  • βœ… Remain when metadata is stripped

Supported Formats

Images: JPEG, PNG, WebP, GIF, TIFF, AVIF, HEIC
Video: MP4, MOV

API Response

{
  success: boolean;
  hasCredentials: boolean;
  
  // Embedded C2PA data
  manifestData?: {
    whoThisComesFrom?: {
      linkedInIdentity?: { name, profileUrl, verified }
      otherIdentities?: [{ name, socialAccounts }]
    };
    aboutThisContent?: {
      actions?: [{ action, softwareAgent, when }]
      genAIInfo?: { generative, training, model }
    };
    aboutTheseCredentials?: { claimSigner, timestamp };
    validationInfo?: { certificate, trustInfo };
  };
  
  // Watermark data (if no embedded found)
  trustMarkData?: {
    identifier: string;    // Watermark payload
    schema: string;        // BCH_SUPER, BCH_5, etc.
    manifestUrl?: string;  // URL to full manifest
  };
  
  error?: string;
}

Filesystem Access

Claude can browse these directories automatically:

  • ~/Desktop

  • ~/Downloads

  • ~/Documents

  • ~/Pictures

No need to provide exact paths! Just ask:

  • "What images are in my Desktop?"

  • "Check recent downloads"

Development

npm run build        # Compile TypeScript
npm run dev          # Development mode
npm run lint         # Check code quality
npm run test         # Run tests
npm run precommit    # Full quality check

Architecture

mcp-content-credentials/
β”œβ”€β”€ src/
β”‚   β”œβ”€β”€ index.ts              # MCP server + filesystem access
β”‚   β”œβ”€β”€ c2pa-service.ts       # Detection orchestration
β”‚   β”œβ”€β”€ trustmark-service.ts  # Watermark detection (Python)
β”‚   β”œβ”€β”€ parsers/              # Data formatters
β”‚   └── types/                # TypeScript definitions
β”œβ”€β”€ scripts/
β”‚   β”œβ”€β”€ install-trustmark.cjs # Auto-installer
β”‚   └── trustmark-decode.py   # Python watermark decoder
└── build/                    # Compiled output

Troubleshooting

"Unable to access that file"

  1. Restart Claude Desktop (most common fix!)

  2. Use absolute paths: /Users/you/... not ~/...

  3. Verify MCP is connected: Ask "What tools do you have?"

"c2patool: command not found"

brew install contentauth/tools/c2patool  # macOS
# or
npm run install-deps

"Python or TrustMark not found"

pip3 install trustmark Pillow
# or
npm run install-deps

No Content Credentials Found

This is normal! The file either:

  • Wasn't created with content authentication

  • Had credentials removed

  • Is a screenshot/copy without provenance

Performance

  • Embedded check: ~150ms (fast path, 80% of cases)

  • + Watermark check: ~600ms (fallback, 20% of cases)

  • First watermark: ~30s (downloads ONNX model, one-time)

Security

  • βœ… Read-only filesystem access

  • βœ… Limited to user directories (Desktop, Downloads, etc.)

  • βœ… Input validation on all paths and URLs

  • βœ… Temporary files auto-deleted

  • βœ… No access to hidden/system files

Contributing

See CONTRIBUTING.md

Resources

License

MIT

Available Tools

2 tools
read_credentials_fileA

Read Content Credentials from a local file. USE THIS TOOL when the user drops a file or provides a file path AND asks questions like: "who made this", "how was this made", "where does this come from", "is this AI", "is this real", "does this have Content Credentials", "what are the c2pa details", or mentions "c2pa" or "Content Credentials". If the user asks a SPECIFIC question (e.g., "Is this AI?"), answer their question directly using the relevant data, then offer to share more details from the Content Credentials. If the user asks GENERALLY about the file or its credentials, present the information in this order (SKIP sections that have no data): 1) Who this comes from (use BULLET POINTS, prioritize LinkedIn verified identities at the top - OMIT THIS ENTIRE SECTION if no CAWG or personal identity is found), 2) About this content (actions taken - EXCLUDE "c2pa.opened" actions unless specifically requested), 3) About these Content Credentials (signer and timestamp), 4) Validation info (certificate details).

Inputs: filePath (required).

read_credentials_urlA

Read Content Credentials from a file at a URL. USE THIS TOOL when the user provides a URL AND asks questions like: "who made this", "how was this made", "where does this come from", "is this AI", "is this real", "does this have Content Credentials", "what are the c2pa details", or mentions "c2pa" or "Content Credentials". Downloads the file temporarily, checks for embedded C2PA manifests, then cleans up. If the user asks a SPECIFIC question (e.g., "Is this AI?"), answer their question directly using the relevant data, then offer to share more details from the Content Credentials. If the user asks GENERALLY about the file or its credentials, present the information in this order (SKIP sections that have no data): 1) Who this comes from (use BULLET POINTS, prioritize LinkedIn verified identities at the top - OMIT THIS ENTIRE SECTION if no CAWG or personal identity is found), 2) About this content (actions taken - EXCLUDE "c2pa.opened" actions unless specifically requested), 3) About these Content Credentials (signer and timestamp), 4) Validation info (certificate details).

Inputs: url (required).

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.1/5.0

    Scored across 2 tools

    Disambiguation5/5

    The two tools have clearly distinct purposes based on input source: read_credentials_file handles local files while read_credentials_url handles URLs. Their descriptions explicitly specify when to use each tool, eliminating any potential confusion about which to select for a given scenario.

    Naming Consistency5/5

    Both tools follow an identical verb_noun pattern (read_credentials_file and read_credentials_url) with perfect consistency. The naming clearly indicates the action (read) and distinguishes the input type (file vs URL) in a predictable manner.

    Tool Count3/5

    With only 2 tools, the server feels somewhat thin for a Content Credentials domain that could benefit from additional operations like validation, manifest listing, or credential creation. While the two tools cover the core reading functionality from different sources, the scope seems limited compared to what a comprehensive C2PA/credentials server might offer.

    Completeness2/5

    The server only provides read operations from two input sources, missing essential capabilities for a Content Credentials domain. There are no tools for creating, updating, validating, or managing credentials, nor tools to list available manifests or check credential status. This creates significant gaps that would limit agent workflows beyond basic information retrieval.

    Maintenance

    ActivityInactive
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      A
      quality
      B
      maintenance
      Verifies C2PA Content Credentials for local files or URLs and returns an LLM-ready verdict on trust, AI generation, and provenance.
      4
      26 npm
      2
      Apache 2.0
    • A
      license
      A
      quality
      C
      maintenance
      Provides MCP-compatible AI agents with a verify_image tool to check a photo's capture time and provenance using C2PA, EXIF, XMP metadata, and pixel forensics, returning a verdict and confidence score.
      1
      74 npm
      2
      MIT
    • A
      license
      A
      quality
      B
      maintenance
      Enables AI agents to analyze images from a URL, local path, or Base64 string and determine whether they are AI-generated, including detection of signatures from Flux, Midjourney, DALLΒ·E, and Stable Diffusion. Also inspects cryptographic C2PA provenance manifests, signing certificate authorities, and camera EXIF metadata without requiring API keys or accounts.
      2
      66 npm
      MIT