Skip to main content
Cognisafe sits between your agent orchestrator and every LLM, MCP server, and tool registry it calls — intercepting every agent action, tool invocation, and inter-agent message at the orchestration layer, then scoring it asynchronously against the full OWASP LLM Top 10. Your app sees no added latency. Your security team gets a complete audit trail, automated threat detection, compliance evidence, and per-agent attribution across the entire agentic pipeline. There is no framework lock-in. Cognisafe works via a lightweight SDK that rewrites your provider’s base URL (proxy mode) or wraps the client’s create method directly. Your existing LLM and agent code stays unchanged beyond a one-line patch call. Works with LangGraph, CrewAI, AutoGen, Semantic Kernel, MCP servers, and any OpenAI-compatible orchestrator. Cognisafe is also the scoring and governance destination for telemetry you already collect: if your agents emit events via Observra or an OpenTelemetry pipeline with GenAI spans, route them to /ingest/* and get the same OWASP scoring, dashboards, and audit evidence — with no SDK change at all.

​
Core components

Runtime proxy

A Go reverse proxy that intercepts every agent action, tool call, and inter-agent message at the orchestration layer — observe mode (zero latency) or block mode (inline enforcement).

Safety scoring

A Python worker pulls jobs from Redis and runs all 10 OWASP LLM scorers asynchronously (LLM-as-judge) — no latency on the hot path. Supports built-in scorers, custom LLM-as-judge prompts, regex, and keyword lists.

Governance dashboard

Request volume, cost, latency, agent-level attribution, safety scores, red team campaigns, RBAC, SOC 2 evidence packs, and one-click compliance PDF export.

SDKs

Python and Node.js SDKs that patch OpenAI, Anthropic, Mistral, and Cohere clients in a single call. Java SDK in beta.

​
Get started

Quickstart

Up and running in under 5 minutes — install the SDK, patch your provider, and see your first request in the dashboard.

How it works

Detailed walkthrough of the proxy, async scoring pipeline, and data model.

​
Supported LLM providers

​
Cloud providers

​
Self-hosted models

​
Supported agent frameworks

Cognisafe is framework-agnostic. Any framework that calls a supported LLM provider is automatically observed once you patch the provider client.
  • LangGraph — multi-agent pipelines, full node-level attribution
  • CrewAI — agent crews, task attribution per crew member
  • AutoGen — conversational multi-agent workflows
  • Semantic Kernel — .NET and Python, Azure-native deployments
  • MCP (Model Context Protocol) — tool call monitoring, server governance
  • Pydantic AI — type-safe agent workflows
  • LlamaIndex — RAG pipelines and data agents
  • OpenClaw / ZeroClaw — autonomous agents with tool access

​
Key capabilities

OWASP LLM Top 10

All 10 OWASP LLM threat categories scored across every request, tool call, and inter-agent message. Severity rated 1–5 with agent-level attribution.

Automated red team

On-demand and scheduled red team campaigns using PyRIT with TAP — tests jailbreaks, PII leakage, system prompt exfiltration, tool abuse, and excessive agency.

Custom scorers

Write your own LLM-as-judge scoring prompts, regex patterns, or keyword lists and attach them to any project.

Compliance evidence

SOC 2 evidence pack, NIST AI RMF mappings, tamper-evident pgaudit trail, and one-click PDF export for audits.