Silicon Transfer MCP Server

FTP / FTPS / SFTP for Claude and other MCP clients - with built-in transfer proofs.

Every upload and download returns a verdict: PROVEN only when the SHA-256 of the remote file matches your local file. No more "transfer complete" on blind trust.

Philosophy: prove it, then claim it. Built from scratch by Silicon Networks - the makers of AEGIS Shield.

Why this one?

Tools (13)

Connection: silicon_connect, silicon_disconnect, silicon_status, silicon_server_info Files: silicon_upload_file, silicon_download_file, silicon_delete_file, silicon_file_info Directories: silicon_list_dir, silicon_make_dir, silicon_upload_dir, silicon_download_dir, silicon_rename

Transfers accept proof: "hash" | "size" | "none" - hash is the default for single files, size for directory trees (switch to hash when it matters).

Install

npm install -g silicon-transfer-mcp-server

That's it - live on npm. Or run it without installing via npx (see config below).

git clone https://github.com/SiliconAINetworks/silicon-transfer-mcp-server.git
cd silicon-transfer-mcp-server
npm install
npm run build

Claude Desktop configuration

Add to your claude_desktop_config.json (no credentials in the config - you pass them at runtime through silicon_connect, for your own server):

{
  "mcpServers": {
    "silicon-transfer": {
      "command": "npx",
      "args": ["-y", "silicon-transfer-mcp-server"]
    }
  }
}

{
  "mcpServers": {
    "silicon-transfer": {
      "command": "node",
      "args": ["C:/path/to/silicon-transfer-mcp-server/dist/index.js"]
    }
  }
}

Usage example

Ask your AI assistant:

"Connect to my server via SFTP (host example.com, user deploy, key at ~/.ssh/id_ed25519) and upload dist/app.js to /var/www/app.js - with hash proof."

The assistant calls silicon_connect, then silicon_upload_file and answers with the verdict:

{
  "proof": {
    "verdict": "PROVEN",
    "detail": "SHA-256 identisch - Transfer bewiesen.",
    "local_sha256": "9f2a...",
    "remote_sha256": "9f2a..."
  }
}

If the hashes ever differ you get MISMATCH with both fingerprints - retry instead of trusting a broken deploy.

Security notes

Built with the OWASP Top 10 for Agentic Applications (2026) in mind - specifically ASI02 (Tool Misuse) and ASI04 (Agentic Supply Chain).

What this server does not do

• No stored credentials. They are runtime-only tool parameters for your own server - no defaults, no config files with secrets, no server addresses shipped.
• No dynamic tool definitions. Every tool is fixed in the source you can read; nothing is fetched from a remote registry at runtime, so no third party can rewrite what your assistant thinks a tool does.
• No telemetry, no phone-home, no analytics.

Sharp edges - operate accordingly

silicon_delete_file, silicon_delete_directory and silicon_sync_deploy (with deleteExtras) can destroy remote data. An assistant reading untrusted content - a web page, a PDF, an email - can be steered by instructions hidden inside it (indirect prompt injection). Treat these tools like rm -rf:

• Connect with an account scoped to the directory you intend to touch, never root.
• Prefer SSH keys over passwords; give the key its own restricted account.
• Run silicon_list_dir before any delete, and read the plan before approving.
• Keep deleteExtras off unless you are deliberately mirroring.

Verifying what you installed

Package name is exactly silicon-transfer-mcp-server, published by aegisshield888. Similar names on npm are placeholders we registered to prevent typosquatting - they contain no code. Source: the GitHub repository linked above.

Reporting

Found a hole? Open an issue on GitHub, or write to security@siliconnetworks.ch.

License

MIT (c) 2026 Silicon Networks. Built from scratch - single-author codebase.