radar MCP Server

Overview

Radar functions as an open-source Kubernetes user interface equipped with a built-in Model Context Protocol server designed for AI agents. Developed in Go under the Apache-2.0 license, the project provides visibility into live cluster data, topology, health statuses, events, and changes. It can run locally as a single binary, as a kubectl plugin, or within a cluster using Helm, operating without requiring an account or cloud dependency.

The platform supplies a shared model of the cluster for both human operators and AI agents, presenting information through a web interface and an MCP endpoint. Users can connect various AI clients to read cluster data safely, while optional write endpoints enable administrative capabilities. The software supports multiple conformant Kubernetes environments and handles large clusters with tens of thousands of pods by watching resources through informers.

What it can do

Radar communicates directly with the Kubernetes API using the user existing kubeconfig configuration. It watches the cluster via informers and pushes real-time updates to the browser using server-sent events. Resource changes are recorded while monitoring the cluster, maintaining history in memory by default with optional SQLite or PostgreSQL persistence across restarts.

The integrated MCP server exposes cluster insights on port 9280 by default, or through a custom port specified via command-line arguments. It offers a read-only endpoint that excludes write operations, alongside a full endpoint containing destructive actions like scaling, restarting, applying configurations, rolling back, triggering GitOps synchronization, and draining nodes.

AI agents query the MCP server to obtain minified context regarding resource topology, health assessments, deduplicated events, filtered logs, and recent changes with diffs. This architecture prevents agents from consuming excessive context window space with raw YAML, allowing them to process incidents and diagnose deployment issues faster than querying raw kubectl directly.

Install and configure

These commands were found in the project's own README. We checked each package name against the public registry.

brew install kubectl-radar

(not a registry package: check the source before running it)

brew install --cask skyhook-io/tap/radar-desktop

(not a registry package: check the source before running it)

It also lists hosted endpoints of type: streamable-http.

Environment variables:

Requirements

The README mentions setup for these clients: Claude Desktop, Claude Code, Cursor, VS Code, Cline, Zed.

Use cases

Engineers can troubleshoot failing workloads by inspecting a ranked queue of broken resources, exploring relationships on the topology view, reviewing change timelines with diffs, and accessing event logs. When an incident occurs, users can hand evidence over to an integrated AI agent for automated investigation and diagnosis.

Teams can manage and monitor workloads across multi-cluster environments, viewing Helm chart releases, native ArgoCD and Flux GitOps synchronization states, traffic patterns, audit logs, and capacity metrics from a unified interface without manually jumping through multiple cluster contexts.

Radar is built for software engineers, platform teams, and on-call operators managing Kubernetes clusters who need a unified visual interface and AI agent assistance for debugging incidents, tracking deployments, and analyzing cluster topology without relying on paid commercial tiers or cloud-synced control planes.

Limitations and things to check

The documentation does not state all hardware resource prerequisites beyond mentioning that the software is tested on clusters containing tens of thousands of pods. Users should verify their kubeconfig permissions, network connectivity to the Kubernetes API server, and ensure that appropriate storage flags are configured if persistent timeline history is required across restarts.

The documentation does not specify every possible client configuration detail or unsupported operating system edge case outside of standard installation paths like Homebrew, binaries, Krew, the desktop app, and Helm charts. Operators must review client-specific MCP configuration requirements before connecting external agents to the read-write or read-only endpoints.

Before you connect it:

• The README mentions: can run commands or code on your machine; reads or writes files; controls or scrapes a web browser; connects to databases; manages infrastructure or cloud resources.
• Read the code or the documentation of the tools it exposes, and note what each tool can do.
• Grant the smallest permissions you can: limit folders, API scopes and database roles.
• Keep credentials in environment variables or a secrets manager, never in a shared config file.
• Run it in a container or a separate user account if it can execute code or touch files.

Release history

The most recent releases published on GitHub:

• v1.19.0 (2026-10-29): ## Radar v1.19.0 Modern Kubernetes visibility. Local-first. No account. No cloud dependency. Fast. Topology, event timeline, and service traffic — plus resource browsing and Helm management. Open source. Free forever. ### Installation Homebrew (macOS/Linux): bash brew install skyhook-io/tap/radar…
• v1.18.4-rc.1 (2026-10-29) (pre-release): ## Radar v1.18.4-rc.1 Modern Kubernetes visibility. Local-first. No account. No cloud dependency. Fast. Topology, event timeline, and service traffic — plus resource browsing and Helm management. Open source. Free forever. ### Installation Homebrew (macOS/Linux): bash brew install…
• v1.18.3 (2026-10-28): ## Radar v1.18.3 Modern Kubernetes visibility. Local-first. No account. No cloud dependency. Fast. Topology, event timeline, and service traffic — plus resource browsing and Helm management. Open source. Free forever. ### Installation Homebrew (macOS/Linux): bash brew install skyhook-io/tap/radar…

Maintenance and community

The repository has 75 contributors, 1,932 commits in the last 12 months, 126 open issues and pull requests. It is maintained by Skyhook, a GitHub organisation with 37 public repositories and 34 followers. GitHub's community checks found a security policy, contribution guidelines, a code of conduct.

Frequently asked questions

Is radar free to use?

The repository is released under the Apache-2.0 license.

When was radar last updated?

The last commit was on 2026-10-31, and the project is marked active by our 12-month check.

Which AI clients does radar work with?

Its README gives setup notes for Claude Desktop, Claude Code, Cursor, VS Code, Cline, Zed. Other MCP-compatible clients may also work, but the project does not say so.

What is Radar?

Radar is an open-source Kubernetes user interface and built-in Model Context Protocol server that lets humans and AI agents inspect cluster topology, events, Helm releases, and GitOps sync states.

How do AI agents connect to Radar?

AI agents connect via the built-in Model Context Protocol server using HTTP endpoints such as the read-only URL running locally on port 9280.

Does Radar store cluster data in the cloud?

No, Radar runs locally on your machine or inside your cluster, keeping all cluster data private unless you explicitly opt in to anonymous usage reporting.

What write actions are available through the MCP server?

The full MCP endpoint provides write tools including restart, scale, apply, rollback, GitOps sync, and node drain, which are marked as destructive for client confirmation.

Sources

This page combines data from:

• radar on GitHub
• Radar | The missing open-source Kubernetes UI
• Radar documentation - Radar
• Kubernetes AI Agent Benchmarks · Radar
• Official MCP Registry entry (registry.modelcontextprotocol.io)