convez.ai
MCP Server
MukundaKatta

llm-output-sanitizer-mcp

MukundaKatta/llm-output-sanitizer-mcp

Ver no GitHub
InicialPluginScore · de 0 a 100

MCP server for @mukundakatta/llm-output-sanitizer: strip dangerous HTML, SQL, and shell snippets from LLM output before it reaches a render, query, or shell sink.

Instalação

$npx convez-ai MukundaKatta/llm-output-sanitizer-mcp

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 2 meses
Versão
—
Issues abertas
4
Contribuidores
1

README

A tradução para o português deste README ainda está na fila.

llm-output-sanitizer-mcp

npm mcp registry

MCP server that exposes @mukundakatta/llm-output-sanitizer to any MCP-aware client (Claude Desktop, Cursor, Cline, Windsurf, Zed).

Strip dangerous HTML, SQL, and shell snippets from LLM output before they reach a render path, a query engine, or a shell.

Tools

NameWhat it does
sanitize_outputReturn cleaned text plus the list of removed patterns.
assert_safe_outputSame scan, but ok: false when anything dangerous was found.
list_findingsEnumerate matches without rewriting the text.

Detectors target the high-leverage sinks: <script>, <iframe>, <form>, <meta>, DROP/TRUNCATE/ALTER, rm -rf, curl ... |, sudo, etc.

Install

// claude_desktop_config.json (Claude Desktop)
// or the equivalent in Cursor / Cline / Windsurf / Zed
{
  "mcpServers": {
    "llm-output-sanitizer": {
      "command": "npx",
      "args": ["-y", "@mukundakatta/llm-output-sanitizer-mcp"]
    }
  }
}

Restart your client. The three tools appear in the tool drawer.

Example

> sanitize_output on "Hello <script>steal()</script> world"

{
  "safe": false,
  "text": "Hello [removed:html] world",
  "findings": [
    { "kind": "html", "match": "<script>" },
    { "kind": "html", "match": "</script>" }
  ]
}

Pass sink: "html" to additionally HTML-escape < > & after stripping, so the cleaned text is safe to drop straight into a DOM.

Why this exists

Even a well-aligned model will repeat dangerous strings it found in retrieved context. A small, deterministic post-filter on the way out is cheap insurance, and it pairs naturally with prompt-injection-shield-mcp on the way in.

This server is a thin wrapper. The detection logic lives in the underlying library and is zero-dependency, sub-millisecond, and entirely local.

License

MIT © Mukunda Katta

Ver o original no GitHub

Bom para

Relacionados