This MCP server classifies prompt injection, jailbreaks, and out-of-scope user input so AI agents can check text safety and scope before acting on it. It exposes a single MCP tool that evaluates untrusted user content for agent use.

🛠️ Key Features

• Provides one tool: classify_input
• Classifies prompt injection, jailbreaks, and out-of-scope user input
• Intended for pre-action safety checks in agents

🚀 Use Cases

• Guardrail step before an agent processes user-provided text
• Filtering or blocking unsafe and out-of-scope requests
• Supporting AI tools with an “AI firewall” classification step

⚡ Developer Benefits

• Simple integration via an MCP tool (classify_input)
• Works with multiple providers (e.g., OpenAI, Anthropic, Google, DeepSeek, Ollama; Google default)
• Configuration via environment variables (e.g., KOMA_PROVIDER, GEMINI_API_KEY)

⚠️ Limitations

• Exposes a single classification tool (classify_input) rather than multiple agent actions

Topics

• ai-security, guardrails, middleware, nodejs, prompt-injection, rag, rate-limiting, typescript, ai-firewall, llm-security, owasp-top-10-llm, ai-governance, ai-tools