A HelmA package manager for Kubernetes. A Helm chart bundles resource templates and default values so you can configure and install an application. values file can look right and still do nothing. Valid YAML, correct indentation, wrong key.
Coding assistants can make this mistake too. Knowing Helm isn’t the same as knowing the chart version you’re using.
I built artifacthub-mcp to give my assistant a way to check. It looks up packages on Artifact Hub, a catalog of cloud native packages, and retrieves chart values, schemas, and templates instead of leaving the model to guess.
It’s a Model Context ProtocolA protocol that lets AI applications connect to servers exposing tools, resources, and prompts. server that runs locally over stdio. Its tools are read-only. It doesn’t install charts or access your cluster.
Check the chart first
The lookup matters more to me than the search. Once the assistant finds a chart, it can fetch the default values.yaml and, when available, its values schema before suggesting changes.
Here’s a prompt to try:
Find prometheus-community/prometheus on Artifact Hub. Use the same chart version for every lookup. Read its default values and any available values schema, then write a values file that enables persistent storage for the Prometheus server. State the version and explain each setting. Do not deploy it.
This isn’t a recorded demo. I’d want to see the assistant fetch the defaults and schema for that version before trusting its answer.
If a setting isn’t clear from the defaults, the server can list the chart’s templates and retrieve their source. That lets the assistant follow a value into the template that uses it, rather than guess what it does.
There’s a limit to this. Reading a template isn’t rendering it. Conditions, dependencies, and cluster capabilities can change the result. I’d still run helm template with the same chart version and my values file, then review the manifests before deploying. Even that doesn’t prove the cluster will accept them.
Check upgrades too
For an existing deployment, the assistant can look up available versions and compare the published changelog for an upgrade. Keep the version explicit. The latest defaults may not match what you’re running, and a changelog isn’t a full diff.
The server also retrieves Artifact Hub’s security reports. Those are published reports, not fresh scans of your deployment. A missing report isn’t a clean one.
Try it
Install with Homebrew or Cargo:
brew tap luxass/tap
brew install luxass/tap/artifacthub-mcpcargo install --locked artifacthub-mcpOr download a pre-built binary for Linux or macOS from GitHub Releases and put it on your PATH.
Check that artifacthub-mcp --version works in your shell, then register it with your client:
claude mcp add --transport stdio artifacthub -- artifacthub-mcpcodex mcp add artifacthub -- artifacthub-mcpClaude Code defaults to local scope for the current project. Add --scope user before the server name to make it available across projects. Codex stores this registration in ~/.codex/config.toml by default. If the client can’t find the binary, replace artifacthub-mcp in the command with its absolute path.
The docs cover other clients, tool filtering, and the full tool list.
The assistant can still get the configuration wrong. I want it to show which version it checked and where each setting comes from. That gives me something to review besides plausible-looking YAML.