docs

MCP Runtime

MCP Runtime is an open-source, Kubernetes-native control plane for deploying, governing, and brokering MCP servers. It packages server deployment, registry workflows, gateway routing, access policy, audit evidence, and observability into one operating surface for platform, security, and compliance teams.

Unlike public MCP directories or client-specific catalogs, MCP Runtime is not just a place to discover servers, and it is not a marketplace for MCP listings. The platform control surface is the front door to a deployable runtime: Kubernetes reconciliation, registry workflow, brokered tool calls, access grants, consented sessions, audit, compliance evidence, and operational visibility. The hosted platform shows what that experience looks like; companies can run the same model inside their own clusters for agents, IDEs, and direct human workflows.

Vendor-neutral MCP infrastructure for platform teams

Build and publish MCP server images, reconcile them with Kubernetes CRDs, expose them through governed gateway routes, and keep policy decisions, consented sessions, audit trails, and telemetry attached to every agent call.

Get started Architecture API reference

Core surfaces

• Operator and MCPServer, MCPAccessGrant, and MCPAgentSession CRDs
• Registry-backed image build, push, and deploy flow
• Sentinel gateway policy, grants, consented sessions, audit, and analytics
• Governance controls for tool access, trust levels, session revocation, and policy versioning
• Compliance-oriented event records for who called what, when, against which server, and whether it was allowed or denied
• Ingress routing for path-based MCP endpoints
• CLI for setup, status, registry, access, Sentinel, and servers

What MCP Runtime installs

mcp-runtime setup installs the CRDs, runtime namespaces, an operator, registry integration, ingress wiring, and the bundled Sentinel stack. Sentinel includes the gateway request path, grant/session policy materialization, analytics ingest and processing, dashboard/API services, and observability components.

Compared with MCP directories

Top MCP directories and catalogs such as Glama, Smithery, Docker MCP Catalog, PulseMCP, mcp.so, and client-specific catalogs are useful for public discovery, metadata, install snippets, or client onboarding. MCP Runtime is different: it is an open-source control plane for operating governed MCP servers inside a company environment.

Others usually provide	MCP Runtime provides
Public discovery and categories	Deployable runtime plus an internal server view when teams need one
Install snippets and connection docs	Kubernetes MCPServer reconciliation and routes
Popularity or metadata signals	Trust, grants, sessions, policy decisions, audit, and compliance evidence
Hosted directory or client-specific UX	Self-hosted, vendor-neutral Kubernetes control plane

As of April 2026, we have not found another open-source MCP product that combines a deployable Kubernetes operator, registry workflow, brokered request path, access/session model, audit pipeline, and operational control surface in one system.

Governance, audit, and compliance

MCP Runtime keeps governance on the live request path instead of leaving it as out-of-band documentation. The gateway evaluates MCPAccessGrant and MCPAgentSession policy before tool calls reach a server, including tool-level allow/deny rules, trust requirements, consented trust, expiry, and revocation.

Each decision can emit audit and analytics events with the server, namespace, human ID, agent ID, session ID, tool name, policy version, decision, reason, and trust context. That gives platform and security teams a queryable record for reviewing access, investigating denied calls, and preparing compliance evidence for governed agent workflows.

Before setup

MCP Runtime expects an already-running Kubernetes cluster and a workstation with the CLI prerequisites installed. The setup flow applies the runtime manifests, installs the operator and Sentinel services, and wires ingress and registry resources for the selected environment.

For provider-specific prerequisites such as container runtime registry trust, DNS, ingress, TLS, and k3s configuration, start with Cluster readiness.

Choose a path

Start Install the platform Build the CLI, run preflight checks, install the stack, and deploy the first server. Understand Read the architecture Trace how the control plane, registry, broker, operator, and Sentinel services fit together. Prepare Check your cluster Review prerequisites for k3s, kind, minikube, Docker Desktop Kubernetes, kubeadm, and EKS. Ship Publish an MCP server Write a manifest or `.mcp` metadata, push an image, deploy it, and verify what the platform creates.

Operate MCP Runtime

Runtime Control plane CRDs, reconciliation outputs, image resolution, ingress wiring, and rollout flow. Governance Sentinel request path Gateway policy, grant/session evaluation, analytics, audit events, and dashboard services. CLI Command reference Setup, status, registry, server, access, pipeline, and Sentinel commands. Reference API and CRDs MCPServer, access grants, sessions, gateway headers, and HTTP APIs. Contribute Read the internals If you are here to contribute, start with the internal docs for codebase structure, package tours, and implementation details.

Common workflows

Workflow	Start here
Evaluate MCP Runtime for a private MCP platform	Getting started, then Architecture
Run MCP Runtime on a real cluster	Cluster readiness, then Runtime
Govern tools and sessions	Sentinel, then API reference
Integrate from automation	CLI, then API reference
Work on the codebase	Internals

Project status

MCP Runtime is alpha. The architecture is stable enough to evaluate as governed MCP infrastructure, but API and UX details are still evolving. Treat the v1alpha1 types as the source of truth.