Skip to content
DeepintShield DeepintShield

v1.0.0 - Initial release

The first release of DeepintShield - one API that fronts every model with real-time guardrails, cost optimization, agentic tool-call authorization, and full observability.

Highlights

Section titled “Highlights”

Unified AI gateway

  • One OpenAI-compatible endpoint to 22+ providers and 2500+ models - OpenAI, Anthropic, Google, Bedrock, and more.
  • Drop-in SDK compatibility: point OpenAI, Anthropic, Google GenAI, Bedrock, LangChain, LiteLLM, or PydanticAI at DeepintShield unchanged.
  • Automatic fallbacks, retries, load balancing, and conditional (rule-based) routing.

Real-time AI guardrails

  • Prompt-injection, jailbreak, toxicity, PII, and secret detection across input, output, RAG, and tool stages - with sub-5ms latency on clean traffic.
  • One-click OWASP LLM Top 10 cards and industry domain packs.
  • Five decision outcomes: allow, redact, require approval, sandbox, deny.

PII redaction & data protection

  • Inline detection and masking of PII, PHI, and secrets at the request, response, and tool boundaries.

RAG security

  • Per-chunk allow / redact / reject / quarantine with source trust scoring, citation enforcement, and one-line retriever/embedder guards in the SDK.

Agentic security (PEP/PDP)

  • Authorize every agent tool-call (allow / deny / require approval / mask) with a visual policy builder, agent identity, tool integrity checks, AIBOM export, and a tamper-evident decision audit. Roll out in Shadow, Canary, or Enforce mode.

Cost optimization

  • Semantic caching, provider prompt caching, cascade routing, request coalescing, prompt compression, and reasoning-effort throttling - cutting LLM spend by up to 90%.

MCP gateway

  • Broker Model Context Protocol servers, group and scope tools, per-tool auth, and a sandboxed code mode.

Governance & accuracy

  • Virtual keys with budgets, rate limits, model allow-lists, and guardrail/cache policy; hallucination defense and response-consistency controls.

Observability

  • Real-time dashboards, per-request tracing, OpenTelemetry export, and agentic traces with scoring.

SDKs

  • Unified Python SDK with one-line tool-gating binders for LangGraph, CrewAI, OpenAI Agents, LlamaIndex, AutoGen, and PydanticAI.

Deployment

Section titled “Deployment”