Kosmoy

Kosmoy is a self-hosted AI management platform that gives enterprises one control plane over every model, application, agent and MCP server they run. It combines an AI inventory, an OpenAI-compatible gateway with guardrails and budgets, observability with cost tracking, and Kubernetes-native sandboxes for autonomous agents — aimed at CIOs who need EU AI Act, ISO 42001 and NIST AI RMF evidence without sending data to a vendor's cloud.

Kosmoy, founded in 2024 and headquartered in Milan with a fully remote team across New York, Buenos Aires, Belgrade, Dubai and Singapore, sells an AI management platform structured as four concentric layers of control. AI Inventory maintains registries of AI systems, models, agents and MCP servers with ownership and risk classification, and pulls agents automatically from Azure AI Foundry, AWS Bedrock, Google Vertex, Salesforce and ServiceNow. Monitoring and Observability tracks cost and usage per model, application and user with budgets that warn and then block, plus roughly twenty evaluators and red teaming. Governance centres on an AI Gateway exposed through an OpenAI-compatible API that authenticates requests, inspects traffic, applies guardrails, routes to approved models, enforces rate limits and logs every decision. AI Action Control adds Action Capsules — Kubernetes-native containers with in-container sandboxing that hold autonomous agents accessing systems of record — supervised by Mission Control with pre-flight authorisation, just-in-time credentials and a kill switch; an Agent Access Control registry giving each agent a distinct identity with approve, review and block states shipped in June 2026. The platform is deliberately not SaaS: it installs via Helm charts into the customer's own Kubernetes on Azure, AWS, GCP or on-premises, is single-tenant by default, air-gap capable and model-agnostic across OpenAI, Anthropic, Google, Meta, Mistral, Hugging Face and private fine-tuned models. It is built to produce the evidence the EU AI Act (Articles 9, 11, 12, 14, 17 and 50), ISO/IEC 42001 and the NIST AI Risk Management Framework require. Kosmoy says it runs in production at Banca d'Italia and Leonardo, and was named a Representative Vendor in Gartner's February 2026 Market Guide for AI Gateways.

Category

Governance & Security

Pricing

Contact for pricing, Subscription

Target Market

CIOs, CTOs, Heads of AI Platform, Chief Risk Officers, Enterprise Architects

Deployment

Self-hosted, Multi-cloud, Hybrid

Founded

2024

Headquarters

Milan, Italy

• ✓
  OpenAI-compatible AI Gateway

  Authenticates, inspects, guardrails, routes, rate-limits and logs every LLM, MCP and agent-to-agent call through one enforcement point.

• ✓
  AI inventory and agent registry

  Registers models, agents and MCP servers with ownership and risk class, pulling agents automatically from Bedrock, Vertex, Foundry, Salesforce and ServiceNow.

• ✓
  Action Capsules

  Kubernetes-native containers with in-container sandboxing that isolate autonomous agents touching systems of record, using just-in-time credentials.

• ✓
  Mission Control kill switch

  Supervises the agent fleet with pre-flight authorisation and an immediate kill switch, which is the control auditors ask for on autonomous agents.

• ✓
  AI FinOps budgets

  Tracks cost and usage per model, application and user, with budgets that first warn and then hard-block further spend.

• ✓
  Compliance evidence generation

  Produces the artefacts the EU AI Act Articles 9, 11, 12, 14, 17 and 50, ISO/IEC 42001 and NIST AI RMF require from deployers.

• ✓
  RAG-in-a-Box

  Prebuilt ingestion pipelines into SharePoint, Confluence and object stores with hybrid retrieval and re-ranking, so retrieval is governed too.

• •
  Multi-cloud AI governance

  An enterprise running models on Azure, AWS and Vertex enforces one guardrail, routing and logging policy across all three from a single control plane.

• •
  EU AI Act audit readiness

  Risk classification, conversation logging and system inventory produce the deployer evidence Articles 9 through 50 require, without a manual documentation project.

• •
  Containing autonomous agents

  Agents reaching into SAP, Salesforce or ServiceNow run inside Action Capsules with scoped just-in-time credentials and a supervisor kill switch.

• •
  Controlling runaway AI spend

  Per-model and per-user budgets warn then block, and the gateway routes cheaper workloads to smaller models to cut cost.

• •
  Air-gapped AI deployment

  Defence and central-banking environments run the full control plane single-tenant inside their own Kubernetes with no external dependency.

Kosmoy publishes no price. It sells an annual enterprise licence quoted against four explicit variables — which control layers you take (inventory, observability, gateway policy, agent containment), workload scope, deployment boundary (Azure, AWS, GCP, on-premises or air-gapped) and implementation support — after a five-step cycle ending in a proof of concept. Third-party directories list a free trial but no free version; budget separately for the Kubernetes infrastructure it runs on.

This page was written from 5 sources, 2 on domains other than kosmoy.com.

1. 1.kosmoy.com — platformvendor
2. 2.kosmoy.com — companyvendor
3. 3.kosmoy.com — pricingvendor
4. 4.sourceforge.net — Kosmoy
5. 5.slashdot.org — Kosmoy