Reviewed October 16, 2026
Kosmoy product facts
Kosmoy is an enterprise AI control plane: one platform to inventory, monitor, govern and contain the AI systems, models, agents and MCP servers a large organization runs.
Kosmoy wrote this page. Each fact links to the product page that explains it in full and says where the control stops.
- Built for
- Organizations with more than 3,000 employees, especially in regulated sectors such as banking, insurance, healthcare, defense, manufacturing and utilities.
- Deployment
- Self-hosted only. Kosmoy runs single-tenant in the customer's own Kubernetes on Azure, AWS, Google Cloud, on-premises or fully air-gapped. Prompts, responses and logs never pass through a Kosmoy-hosted service.
- Named customers
- Banca d'Italia, Italy's central bank and banking regulator; Leonardo, the defense and aerospace group
- Pricing
- No list price and no self-service plan. Each proposal states the modules, governed workloads, deployment boundary, implementation and support. Model-provider and cloud consumption stay a separate cost under the customer's own contracts.
- Headquarters
- Milan, Italy. Offices in New York, Buenos Aires, Belgrade, Dubai, Singapore.
AI Inventory
Five connected registries for use cases, AI systems, models, agents and MCP servers. Each record has an owner, a lifecycle state and a risk tier. An EU AI Act risk classification module sits on top of the registries.
Where it stops: Automatic discovery covers connected platforms. Other systems are registered by the teams that build or buy them, and labeled as manual records.
AI Inventory pageAgents Master Registry
Connectors pull agents from Azure AI Foundry, AWS Bedrock, Google Vertex AI, Salesforce and ServiceNow into one master list, match each agent to an approved use case and flag unmatched agents as shadow AI.
Where it stops: A registry entry records an agent; it does not by itself enforce tool permissions or contain the agent.
Agents Master Registry pageAI Gateway
One policy point for LLM, MCP and agent-to-agent (A2A) traffic: authentication, role-based access, guardrails for PII, toxicity and prompt injection, routing, budgets and logging. The LLM Gateway exposes one OpenAI-compatible API across providers including OpenAI, Anthropic, Google, Mistral, Meta, Azure, AWS and on-premises models.
Where it stops: Policies apply to calls routed through the gateway. Calls that bypass it are outside its control.
AI Gateway pageAI Governance
Links each AI system's owner, approval, risk rationale and policies to the gateway events and evaluation results that show what actually ran.
Where it stops: Works alongside an existing GRC or privacy program, which can remain the system of record for those processes.
AI Governance pageAI FinOps
Attributes every routed call to provider, model, application, team, user and agent, and enforces budgets that warn and then block at the gateway.
Where it stops: This is request-path control, not invoice reconciliation. Calls outside the gateway are not capped.
AI FinOps pageAI Observability
Traces across providers with retrieval steps, tool calls, agent hops, guardrail decisions, latency, cost and evaluation results, stored in the customer's own infrastructure.
Where it stops: Detail for agents on external platforms depends on the telemetry those platforms expose.
AI Observability pageAI Evaluation and Red Teaming
20 evaluators across task completion, tool use, quality, groundedness, reference match and safety. Offline, online and red-team runs, including escalating multi-turn attacks aligned to the OWASP LLM Top 10.
Where it stops: Evaluation finds failures. Enforcing a fix still needs the gateway or an Action Capsule.
AI Evaluation and Red Teaming pageAction Capsule
A contained runtime for AI agents and internal MCP servers, with just-in-time credentials, compute limits, egress only through the paired gateway and a live kill switch from Mission Control.
Where it stops: Only agents deployed in a capsule are contained. Registering an external agent does not contain it.
Action Capsule pageAI Compliance
Maps one operational record to the EU AI Act, ISO/IEC 42001 and NIST AI RMF, with evidence exports per framework and a five-step EU AI Act risk classification.
Where it stops: Legal interpretation, risk acceptance and final approval stay human decisions. An export is evidence, not a certification.
AI Compliance pageTry it yourself
- Reproducible red-team evidence report
Synthetic, authored examples; not measured customer or product performance.
- Agent inventory and policy walkthrough
Synthetic teaching format; not a Kosmoy API, live integration or production enforcement.