OpenCost MCP server: three tools, already running

A server nobody has to install, what its three tools cover, and why being on by default is the fact that matters most here.

Last verified 1 September 2026 · from OpenCost's own documentation and the `opencost/opencost` repository

This page is one of 90 in a directory of monitoring and developer tools. Each page checks what an AI coding agent can get out of the tool through its MCP server, the connector that lets the agent query the tool directly.

What is the OpenCost MCP server?

It is OpenCost's server for the Model Context Protocol. OpenCost is the CNCF (Cloud Native Computing Foundation) project that measures Kubernetes spend, and it is the upstream of several commercial cost products.

The three tools map onto OpenCost's three data models. Allocation is cost attributed to Kubernetes objects, assets are the underlying infrastructure, and cloud costs are what the provider billed.

Each takes a window parameter as its only required argument, written the way OpenCost writes time windows: 7d, 1h, 30m. get_allocation_costs accepts more. You can aggregate, set a step, choose whether values accumulate, and set two separate controls for idle cost.

OpenCost exposes allocation, asset and cloud-cost queries from the service built into each deployment. That gives an agent Kubernetes financial data, but no code-review, CI or agent-observability surface.

For an agent, that means OpenCost answers what a namespace, workload or node cost over a window. It measures and does not recommend: there are no tools that recommend a new CPU or memory request, and none that name a saving. The agent gets numbers and reasons about them itself.

What can an AI agent do with OpenCost?

The idle switches deserve a decision before the first query. share_idle distributes unused capacity across workloads and include_idle reports it separately, so the same window can produce two defensible totals.

What connecting costs before the first question

A context window is the amount of text a model can hold at once. Three cost tools are built into every OpenCost deployment and enabled by default on port 8081. They cost 496 to 645 tokens on Claude, which is close to the floor for any connection at all.

Because they are on by default, this is context an OpenCost deployment spends whether the team asked for MCP or not.

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 from the 3 tool entries in the vendor's own documentation. That is the only public inventory.

The server could not be run for a live tools/list. So no input schemas are available. These figures cover each tool's name and description only. They are a floor, not the figure.

The Claude columns are measured through the API's count_tokens endpoint, with an empty schema per tool. So they cover the same content as the GPT columns. The range there is two tokenizer generations. 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.

OpenCost MCP server limits: where answers come back incomplete

OpenCost starts by default and reports without interpreting; those properties explain its limits.

How to configure OpenCost MCP for agents

1. 1

   Decide whether you want it on

   It starts automatically with OpenCost on port 8081. The documentation states users have full control to disable it or change the port.

2. 2

   Treat port 8081 as an interface

   It speaks HTTP, so whatever can reach that port in your cluster can call these tools.

3. 3

   Settle the idle question

   share_idle and include_idle change what a cost figure means, and neither is more correct than the other.

4. 4

   Know what is absent

   There are no savings, rightsizing or recommendation tools. OpenCost reports; it does not advise.

Paste this into your agent’s instructions

Two switches change a cost total without changing the query, and the window syntax is OpenCost's own. Both belong in the answer.

When you answer using OpenCost MCP tools, state:
- The window you queried, in OpenCost's own syntax.
- How idle cost was handled. share_idle and include_idle
  produce different totals for the same period.
- Which tool the figure came from. Allocation, asset and
  cloud cost are three different views of spend.
Never state that data does not exist. State that no cost
data was returned for the window 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 reporting. OpenCost has a well-documented HTTP API, and the three tools are thin wrappers over it. For a dashboard or a scheduled export, calling the API directly is simpler.

The connector earns its place in a conversation about a specific workload, where the agent can move from a namespace total to the assets underneath it without leaving the editor.

For savings analysis over the same data, Kubecost builds rightsizing, abandoned-workload detection and quota recommendations on top of OpenCost.

Use the API for reporting. Connect the server when an agent is investigating a specific cost, and reach for Kubecost when you want advice rather than numbers.

OpenCost MCP server setup

There is nothing to install. If OpenCost is running, the MCP server is running with it, on port 8081.

The documentation describes it as zero configuration and built into the official Helm chart for production deployments.

claude mcp add --transport http opencost \
  http://opencost.opencost.svc:8081

{"mcpServers":{"opencost":{
  "url":"http://opencost.opencost.svc:8081"}}}

[mcp_servers.opencost]
url = "http://opencost.opencost.svc:8081"

Every other client

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

# Configure or disable the MCP server through the
# OpenCost Helm chart values, where it is bundled.

kubectl port-forward -n opencost svc/opencost 8081:8081

Read-only access, permissions and security

• It is on unless you turn it off

  The server starts automatically with every OpenCost deployment, which is the opposite default from most connectors here.

• There is no authentication in the server itself

  It listens on port 8081 over HTTP, so network policy is the boundary.

• Every tool reads

  None of the three creates, changes or deletes anything.

• Cost data describes your infrastructure

  Namespace and workload names travel to the model with the numbers, which is a disclosure even though no credential is involved.

Troubleshooting

The endpoint is not reachable

Check the port. The server runs on 8081 by default, which is separate from the OpenCost API port most deployments already expose.

A query is rejected

Have the agent read the tool schema before it composes a call. window is required on all three tools and uses OpenCost's syntax, such as 7d, 1h or 30m.

Two people get different totals for the same period

Compare the idle settings. share_idle distributes idle cost across workloads and include_idle reports it as its own line.

There are no savings recommendations

There are none to find. OpenCost measures cost; Kubecost builds rightsizing and savings analysis on top of the same data.

OpenCost MCP server: Reference

What engineers report

OpenCost is a widely adopted CNCF project, so the adoption figures below describe the project rather than the connector.

Tool inventory, the default-on behavior, the port and the Helm bundling were read from OpenCost's own MCP documentation on 24 August 2026. Repository facts come from the GitHub API on the same date. The server is part of the OpenCost binary, so no standalone tools/list capture was possible.

Should you connect an agent to OpenCost?

Yes. Best for in-cluster use with port 8081 restricted, and the idle settings agreed before anyone quotes a total.

Strongest fit

Teams already running OpenCost who want cost questions answered without adding anything.

Main advantage

Zero installation and zero configuration: it is already running in every deployment.

Main weakness

Three tools and no recommendations, so an agent gets numbers and must interpret them itself.

Operational risk

It listens by default with no authentication, so cluster network policy is the only boundary.

Frequently asked questions