Plugin System & Lifecycle

Naagmani's plugin engine allows developers to intercept, modify, and guard AI requests using an out-of-process, language-independent architecture (naagmani.plugin/v1).

graph LR
    Client --> Hook1["Before Ingest Hook"]
    Hook1 --> Guard["Guardrails & DLP"]
    Guard --> Hook2["Before Dispatch Hook"]
    Hook2 --> LLM["Provider Inference"]
    LLM --> Hook3["After Complete Hook"]
    Hook3 --> Client

Key Capabilities #

1. Multi-Language SDKs: Author plugins in Go, TypeScript/Node.js, or Python.
2. Isolated Sandboxing: Plugins run in isolated worker processes, ensuring custom code cannot crash the core gateway engine.
3. Execution Hooks:
   • before_ingest: Inspect incoming headers, validate custom tokens.
   • before_dispatch: Mask sensitive PII/DLP data, inject system context, RAG vector retrieval.
   • after_complete: Sanitize model responses, log analytics, trigger downstream webhooks.

Next Steps #

• Deep dive into plugin development: Plugin Development Guide
• Plugin manifest specification: Plugin Manifest