Kubecost MCP server: 11 read-only Kubernetes cost tools

A first-party server with eleven read-only tools, and the savings half is the half worth connecting for.

Checked 6 October 2026 · from Kubecost's repository, its README and the server's own tools/list response

One of 91 guides to monitoring and developer tools. Each covers what an AI coding agent can get out of the tool through MCP, the protocol that lets the agent query the tool directly.

What is the Kubecost MCP server?

It is Kubecost's server for the Model Context Protocol. Kubecost allocates Kubernetes spend, so a connected agent reaches cost by cluster, namespace and workload.

The reporting half is two tools. kubecost_get_workload_costs returns allocation grouped by cluster and workload, and kubecost_get_cost_comparison compares two time windows, which is how a spike gets attributed.

The savings half is seven tools, each pointed at a different waste pattern. Oversized containers, oversized node groups and oversized volumes are three of them.

The rest cover volumes nobody claimed, underused local disks, missing resource quotas, and workloads with almost no network traffic.

Kubecost's eleven read-only tools concentrate on Kubernetes allocation, savings opportunities and workload rightsizing. They explain cluster spend and state, without covering source review, releases or agent behavior.

In practice, Kubecost answers what a cluster costs and where the waste is. The savings tools are the reason to connect it: they name specific workloads with specific recommendations. It cannot apply any of them, so the change still goes through your normal path.

What can an AI agent do with Kubecost?

kubecost_get_savings_overview is the right first call. It ranks every savings category, so the agent can go straight to the one worth the most rather than calling all seven savings tools.

What the tool definitions cost in context

A client that loads every tool definition at the start of a session spends this much context before the first question. Claude Code defers definitions by default through tool search, and loads a tool's definition when the agent needs it. The method below gives the version and date measured.

A context window is the amount of text a model can hold at once. A client that loads every tool definition upfront sends each tool's name, description and input schema to the model when the session starts. That list is charged whether or not a single tool is called.

On 25 August 2026, 11 tools cost between 5,140 and 9,055 tokens, which is 2.6% to 4.5% of a 200,000-token window. Later releases renamed every tool, so current figures differ.

The descriptions and the schemas are both large. Descriptions average 841 characters and schemas 1,079, because each savings tool explains its filters and its sizing assumptions.

Encodings: GPT-5.6, GPT-5.5 o200k_base; GPT-4, GPT-3.5 cl100k_base; Claude Haiku 4.5 to Opus 5 count_tokens. Counted 25 August 2026 from tools/list on the package installed from the repository. It was run locally with a placeholder base URL.

Registration is static, so no Kubecost instance answered. Figures cover name, description and input schema. The two models in column one share the o200k_base encoding, so they get one count. The spread is formatting, about 22%. Reasoning tokens are output, spent per request. This count is billed as input whenever the client loads it.

The Claude figures are measured rather than encoded. Anthropic publishes no local tokenizer. The definitions were sent to the API's count_tokens endpoint, and a no-tools baseline was subtracted. The range covers two tokenizer generations rather than two serializations. Haiku 4.5 and Opus 4.6 sit at the low end. Opus 5 and Sonnet 5 sit at the high end. The gap between them is about a third.

Kubecost MCP server limits: where answers come back incomplete

Kubecost returns recommendations, not applied changes. The limits below follow from that.

How to configure Kubecost MCP for agents

1. 1

   Decide between the bundled and the standalone version

   The README says the MCP is bundled with the Kubecost Helm installation by default, and that the repository may carry newer versions.

2. 2

   Check the Kubecost version

   The README states compatibility with any Kubecost 3.x, with release notes worth reading for dependencies.

3. 3

   Set the base URL

   KUBECOST_BASE_URL is required, and the server exits with a named configuration error without it.

4. 4

   Read the auth documentation separately

   Authentication, OIDC, API keys and pod hardening live in the repository's own auth documentation rather than the main README.

Paste this into your agent’s instructions

Every savings number is a recommendation under an assumed sizing profile, and none of it has been applied. The block keeps both facts in the answer.

When you answer using Kubecost MCP tools, state:
- The cluster, namespace and time window the numbers cover.
- That savings figures are recommendations, not applied
  changes, and which sizing profile produced them.
- For abandoned workloads, that the signal is abnormally low
  network traffic rather than a proven lack of use.
Never state that data does not exist. State that no workloads
matched the window and filters above.
Write your answer in ASD-STE100 Simplified Technical
English. Use short sentences with one idea in each.

Do you need the MCP server at all?

Not for a monthly review. Kubecost's own interface shows the same savings, and it costs nothing in context.

The connector earns its place when the agent is already in the cluster's manifests. Reading a rightsizing recommendation and editing the resource block is one conversation rather than two tools.

OpenCost is the upstream open-source project and ships its own MCP server, which covers allocation without Kubecost's savings analysis.

Connect it when an agent is changing manifests. Use the interface for periodic review.

Kubecost MCP server setup

The simplest route is the Helm chart, because the MCP ships bundled with the Kubecost installation by default.

For a local client, install the package from the repository and point KUBECOST_BASE_URL at your Kubecost service. An HTTP entry point is also provided.

claude mcp add-json "kubecost" '{"command":"mcp-kubecost",
  "env":{"KUBECOST_BASE_URL":"http://kubecost.internal:9090"}}'

{"mcpServers":{"kubecost":{"command":"mcp-kubecost",
  "env":{"KUBECOST_BASE_URL":"http://kubecost.internal:9090"}}}}

[mcp_servers.kubecost]
command = "mcp-kubecost"
env = { KUBECOST_BASE_URL = "http://kubecost.internal:9090" }

Every other client

Each block below is the configuration for one client, with the file path and the key that client expects.

# The MCP ships with the Kubecost Helm install by default.
# The repository's own chart carries newer versions:
#   charts/mcp-kubecost

KUBECOST_BASE_URL=http://kubecost.internal:9090 \
mcp-kubecost-http

Read-only access, permissions and security

• Every tool reads

  The README states all eleven are read-only, and the captured list contains no create, update or delete tool.

• Nothing is applied

  Rightsizing and quota tools return recommendations; changing a workload still goes through your normal deployment path.

• Telemetry is flagged as experimental

  The README carries a telemetry section marked experimental, which is worth reading before a production rollout.

• Auth is documented apart from the README

  Authentication, OIDC, API keys and pod hardening live in the repository's auth documentation.

Troubleshooting

The server exits at startup

KUBECOST_BASE_URL is missing. The server raises a configuration error naming that variable and stops before registering tools.

Cost queries reject the time window

Call kubecost_list_windows first. It returns the valid windows a Kubecost cost query accepts.

Rightsizing numbers look aggressive

Container recommendations depend on a sizing profile. The README documents the profiles and the percentile each uses, so the profile is part of the answer.

A workload is flagged abandoned but is in use

kubecost_get_abandoned_workloads finds pods with abnormally low network traffic. A workload that does local work without network traffic will look abandoned to it.

Tools go missing after an upgrade

v0.18.0 renamed every tool and prompt to the kubecost_<verb>_<noun> pattern. Update allowlists and saved prompts that use the old names, such as get_savings_overview.

Kubecost MCP server: Reference

What engineers report

The repository is new and has little adoption to report, so the facts below are its own metadata and the wire capture.

Repository facts from the GitHub API, retrieved 6 October 2026. Tool counts come from the server's own tools/list response on 25 August 2026, and match the README's stated count of eleven. Tool names follow the v0.18.0 source.

Should you connect an agent to Kubecost?

Yes. Best for the bundled Helm deployment, with savings treated as candidates rather than decisions.

Strongest fit

Kubecost users who want savings recommendations named in the same session where the manifests get edited.

Main advantage

Seven savings tools covering distinct waste patterns, ranked by kubecost_get_savings_overview.

Main weakness

New and unproven: created in June 2026 with no adoption signal yet.

Operational risk

Low. Every tool reads, and nothing it recommends is applied automatically.

Frequently asked questions