Quick Start - Gateway
Connect a provider, create a Virtual Key, and make your first guarded LLM call.
DeepIntShield
Real-time guardrails for prompts, agents, and tools - authorize govern tool calls, detect PII and jailbreaks, and measure hallucination risk with configurable inline and post-response controls.
Near-zeroDeterministic local guard path
~2µsLocal cached-verdict benchmark
Up to 90%Published example ceiling
29Built-in provider identities
OWASPLLM + Agentic Top-10
MCPGateway & tool control
Agent IDIdentity & auth
RuntimeAgent discovery & review
ScoredHallucination-risk signals
Start here
Section titled “Start here”Performance & Cost
Five optimizations ship enabled (or one flag away): embedded runtime, speculative dispatch, async post-guards, per-category timeouts, semantic cache short-circuit.
AI Guardrails
Configure safety policies across input, output, action, MCP, and RAG evaluation stages.
Python SDK
Use native provider clients, explicit guardrails, RAG filtering, MCP adapters, framework binders, and Agentic enforcement from one package.
Providers & protocols
Configure the built-in provider catalog, stored Responses, cached content, Realtime transports, and OCR.
Enterprise delivery
Deliver control-plane events through the durable webhook outbox and S3-compatible payload storage.
Benchmark figures above are workload- and environment-specific, not service level guarantees. Review the methodology and run your own workload before setting latency, throughput, or savings targets.
What makes DeepIntShield different
Section titled “What makes DeepIntShield different”Real-time, in-process
In an embedded-runtime deployment, eligible guard evaluation runs in the gateway without a separate guard-runtime RPC. ML detectors can still use a sidecar or separate runtime.
Speculative dispatch
Fire the provider call in parallel with input guards. Allow-path latency becomes
max(guards, model) instead of guards + model.
Smart caching cuts spend
Warm exact cache hits make no provider, semantic, or guard-runtime call and avoid database, network, and embedding work. A guarded semantic hit gets one current validation before its final governed response is promoted for later exact reuse. Cold and external/model-dependent work retains real latency. Some repetitive chatbot examples show up to 60% cost reduction; safe hit rate determines the result.
Audit workflows
Configured persisted decisions and findings feed local evidence and export workflows. Verify retention, redaction, and export delivery for your controls.
Get involved
Section titled “Get involved”GitHub
Star, file issues, or send PRs.
Discord
Real-time help, roadmap discussion, design reviews.