Kubernetes Tools

This package provides a collection of Kubernetes functions to be used by Agents. They can be passed directly to an agent as tools or placed behind an MCP server (included). Some use cases include:

• Chat with your kubernetes cluster via GitHub CoPilot or Cursor.
• Build agents to monitor your cluster or perform root cause analysis.
• Vibe-code a custom chat UI.
• Use in non-agentic automations.
• Snapshot a live cluster to a file and replay it as a test fixture, with no cluster needed — see Capturing and replaying a cluster.

Methodology

Our goal is to focus on quality over quantity -- providing well-documented and strongly typed tools. We believe that this is a critical in enabling agents to make effective use of tools, beyond simple demos.

These are built on top of the kubernetes Python API (https://github.com/kubernetes-client/python). There are three styles of tools provided here:

1. There are tools that mimic the output of kubectl commands (e.g. get_pod_summaries, which is equivalent to kubectl get pods). Strongly-typed Pydantic models are used for the return values of these tools.
2. There are tools that return strongly typed Pydantic models that attempt to match the associated Kubernetes client types (see https://github.com/kubernetes-client/python/tree/master/kubernetes/docs). Lesser used fields may be omitted from these models. An example of this case is get_pod_container_statuses.
3. In some cases we simply call to_dict() on the class returned by the API (defined in https://github.com/kubernetes-client/python/tree/master/kubernetes/client/models). The return type is dict[str,Any], but we document the fields in the function's docstring. get_pod_spec is an example of this type of tool.

Currently, the priority is on functions that do not modify the state of the cluster. We want to focus first on the monitoring / RCA use cases. When we do add tools to address other use cases, they will be kept separate from the read-only tools so you can still build "safe" agents.

Installation

Via pip:

pip install k8stools

Via uv:

uv add k8stools

This installs three commands: k8s-mcp-server (the MCP server), k8s-mcp-client (a client for manual testing), and k8s-capture-state (snapshot a cluster to a replayable file).

What changed in each release is in the changelog.

Current tools

These are the tools we define:

• get_cluster_info - which cluster the tools are answering from: kubeconfig context, API server URL and version, or the capture being replayed
• get_namespaces - get a list of namespaces, like kubectl get namespace
• get_node_summaries - get a list of nodes, like kubectl get nodes -o wide (includes capacity/allocatable/conditions/taints/labels)
• get_pod_summaries - get a list of pods, like kubectl get pods -o wide
• get_pod_container_statuses - return the status for each of the container in a pod
• get_pod_events - return the events for a pod
• get_pod_spec - retrieves the spec for a given pod
• get_logs_for_pod_and_container - retrieves logs from a pod and container (supports tail, since_seconds, and previous). See Previous-instance log semantics for when previous=True returns the same text as previous=False — it is Kubernetes behavior, not a bug.
• get_deployment_summaries - get a list of deployments, like kubectl get deployments
• get_replicaset_summaries - get a deployment's replica sets with their revision numbers and images. A deployment's replica sets are its change history: use this to see when a workload last changed and what the change was.
• get_service_summaries - get a list of services, like kubectl get services (includes selector/labels/annotations)
• get_configmap_summaries - get a list of ConfigMaps, like kubectl get configmaps
• get_configmap - retrieve the full contents of a single ConfigMap
• get_statefulset_summaries - get a list of StatefulSets, like kubectl get statefulsets
• get_cronjob_summaries - get a list of CronJobs, like kubectl get cronjobs
• get_job_summaries - get a list of Jobs, like kubectl get jobs
• get_logs_for_job - retrieve logs from a Job's most-recent pod
• get_logs_for_cronjob - retrieve logs from a CronJob's most-recent run
• get_pvc_summaries - get a list of PersistentVolumeClaims, like kubectl get pvc (resolves mounting pods)
• get_events - list cluster/namespace-wide events with server-side filtering

We also define a set of associated "print_" functions that are helpful in debugging:

• print_namespaces
• print_node_summaries
• print_pod_summaries
• print_pod_container_statuses
• print_pod_events
• print_pod_spec
• print_deployment_summaries
• print_service_summaries
• print_configmap_summaries
• print_configmap
• print_statefulset_summaries
• print_cronjob_summaries
• print_job_summaries
• print_pvc_summaries
• print_events

Using the tools

Directly use in an agent

The core tools are in k8stools.k8s_tools. Here's an example usage in an agent:

from pydantic_ai.agent import Agent
from k8stools.k8s_tools import TOOLS

agent = Agent(
        model="openai:gpt-4.1",
        system_prompt=SYSTEM_PROMPT,
        tools=TOOLS
)

result = agent.run_sync("What is the status of the pods in my cluster?")
print(result.output)

The tools use KUBECONFIG (or ~/.kube/config) and its current context. To pin a cluster instead, call configure before the first tool call:

from k8stools import k8s_tools

k8s_tools.configure(kubeconfig="~/.kube/prod.yaml", context="prod-eu")

See Selecting a cluster for how the selection is resolved.

Using via MCP

The script k8s-mcp-server provides an MCP server for the same set of tools. Here are the command line arguments for the server:

usage: k8s-mcp-server [-h] [--transport {streamable-http,stdio}] [--host HOST] [--port PORT]
                      [--log-level {DEBUG,INFO,WARNING,ERROR,CRITICAL}] [--debug]
                      [--kubeconfig PATH] [--context NAME] [--mock] [--state-file FILE]
                      [--state-time {advancing,frozen}] [--no-redact]

Run the MCP server.

options:
  -h, --help            show this help message and exit
  --transport {streamable-http,stdio}
                        Transport to use for MCP server [default: stdio]
  --host HOST           Hostname for HTTP service [default: 127.0.0.1]
  --port PORT           Port for HTTP service [default: 8000]
  --log-level {DEBUG,INFO,WARNING,ERROR,CRITICAL}
                        Log level [default: INFO]
  --debug               Enable debug mode [default: False]
  --kubeconfig PATH     Kubeconfig file to use [default: KUBECONFIG, then ~/.kube/config]
  --context NAME        Kubeconfig context to use [default: K8STOOLS_CONTEXT, then the
                        kubeconfig's current-context]. A kubeconfig or context that is given but
                        cannot be loaded stops the server rather than falling back to another
                        cluster.
  --mock                Run mock versions of the tools that don't need a cluster
  --state-file FILE     Serve a captured cluster snapshot from FILE (implies --mock)
  --state-time {advancing,frozen}
                        Replay clock for a captured snapshot [default: advancing]
  --no-redact           Disable secret redaction of tool output (on by default)

Selecting a cluster

Each server answers from exactly one cluster, chosen at startup:

With none of these given and no usable kubeconfig, the server falls back to the in-cluster service account, as when it runs in a pod.

The server logs the binding when it starts:

WARNING:root:Bound to cluster: context 'prod-eu' at https://prod-eu.example:6443 (kubeconfig: /home/me/.kube/prod.yaml)

and the get_cluster_info tool reports it to an agent, so an agent can check which cluster it is talking to without leaving MCP.

A few rules keep a server from quietly answering from the wrong cluster:

• An explicit selection never falls back. If a --kubeconfig, --context or K8STOOLS_CONTEXT is given but cannot be loaded (a missing file, a misspelled context), the server exits with an error. It does not fall back to in-cluster config, which in a pod would silently mean that pod's cluster.
• The binding is made once. Every tool shares it, and it does not follow later changes to the kubeconfig: running kubectl config use-context does not repoint a server that is already running.
• --kubeconfig and --context are rejected with --mock or --state-file, which serve a capture rather than a cluster.

To serve two clusters, run two servers:

k8s-mcp-server --transport=streamable-http --port=8001 --context prod-eu
k8s-mcp-server --transport=streamable-http --port=8002 \
    --kubeconfig ~/.kube/eks.yaml --context staging

Use MCP with the stdio transport

The stdio transport is best for use with local Coding Agents, like GitHub CoPilot or Cursor. It is the default, so you can run the k8s-mcp-server script without arguments. Here's an example mcp.json configuration:

{
   "servers": {
      "k8stools-stdio": {
         "command": "${workspaceFolder}/.venv/bin/k8s-mcp-server",
         "args": [
         ],
         "envFile": "${workspaceFolder}/.envrc"
      }
   }
}

This assumes the following:

1. The Python virtual environment is expected to be in .venv under the root of your VSCode workspace
2. You have installed the k8stools package into your workspace
3. The environment file .envrc contains any variables you need defined. In particular, you may need to set KUBECONFIG to point to your kubectl config file, and K8STOOLS_CONTEXT to pin a context (or pass --kubeconfig / --context in args).

Use MCP with the streamable HTTP transport

The streamable http transport is enabled with the command line option --transport=streamable-http. It will start an HTTP server which listens on the specified address and port (defaulting to 127.0.0.1 and 8000, respectively). This transport is best for cases where you want remote access to your MCP server.

Here's a short example that starts the server and then does a sanity test using curl to get the tool information:

# start the server
 $ k8s-mcp-server --transport=streamable-http
[07/21/25 19:55:13] INFO     Starting with 18 tools on transport streamable-http          mcp_server.py:59
INFO:     Started server process [6649]
INFO:     Waiting for application startup.
INFO     StreamableHTTP session manager started         streamable_http_manager.py:111
INFO:     Application startup complete.
INFO:     Uvicorn running on http://127.0.0.1:8000 (Press CTRL+C to quit)

# Now, open another terminal window and test it
$ curl -v \
     -H "Content-Type: application/json" \
     -H "Accept: application/json, text/event-stream" \
     -d '{
           "jsonrpc": "2.0",
           "id": 1,
           "method": "tools/list",
           "params": {}
         }' \
     http://127.0.0.1:8000/mcp
*   Trying 127.0.0.1:8000...
* Connected to 127.0.0.1 (127.0.0.1) port 8000
> POST /mcp HTTP/1.1
> Host: 127.0.0.1:8000
> User-Agent: curl/8.7.1
> Content-Type: application/json
> Accept: application/json, text/event-stream
> Content-Length: 120
>
* upload completely sent off: 120 bytes
< HTTP/1.1 200 OK
< date: Tue, 22 Jul 2025 02:56:25 GMT
< server: uvicorn
< cache-control: no-cache, no-transform
< connection: keep-alive
< content-type: text/event-stream
< x-accel-buffering: no
< Transfer-Encoding: chunked
<
event: message
data: {"jsonrpc":"2.0","id":1,"result":{"tools":[.... long text elided ...]}}

Mock tools

When building agents, it can be helpful to test them against mock versions that do not go against a real cluster, but return realistic values. The module k8stools.mock_tools does just that. The data values were captured when running against a real Minikube instance running the Open Telemetry Demo application. When running the MCP server, this may be enabled by using the --mock command line option.

The mock serves a capture — a JSON snapshot of one cluster — rather than a set of per-tool canned answers, so the tools agree with each other: a pod that is not in the capture does not exist for any tool, and one that is has consistent statuses, events, spec and logs. You can snapshot your own cluster the same way — see Capturing and replaying a cluster.

Capturing and replaying a cluster

k8s-capture-state snapshots a live cluster into a single JSON file, and the MCP server replays that file as if it were the cluster. A whole scenario — a crash loop, a full volume, a bad rollout — becomes a fixture you can re-run, commit, or attach to an issue, with no cluster needed to reproduce it.

# Snapshot the current cluster (respects KUBECONFIG)
k8s-capture-state -o incident-1234.json

# Replay it
k8s-mcp-server --state-file incident-1234.json

k8s-capture-state options

usage: k8s-capture-state [-h] [--namespace NS [NS ...]] [-o FILE]
                         [--kubeconfig PATH] [--context NAME] [--no-logs]
                         [--max-log-lines MAX_LOG_LINES] [--no-previous-logs]
                         [--no-redact]
                         [--log-level {DEBUG,INFO,WARNING,ERROR,CRITICAL}]

options:
  -h, --help            show this help message and exit
  --namespace NS [NS ...]
                        Namespace(s) to capture namespaced resources from
                        [default: all]. Namespaces and nodes are always
                        captured in full.
  -o FILE, --output FILE
                        Output file [default: k8s-state-<timestamp>.json]
  --kubeconfig PATH     Kubeconfig file to capture from [default: KUBECONFIG,
                        then ~/.kube/config]
  --context NAME        Kubeconfig context to capture from [default:
                        K8STOOLS_CONTEXT, then the kubeconfig's current-
                        context]
  --no-logs             Skip container logs
  --max-log-lines MAX_LOG_LINES
                        Log lines to capture per container [default: 1000]
  --no-previous-logs    Skip the previous-instance logs of restarted containers
  --no-redact           Disable secret redaction of captured values (redaction
                        is on by default; can also be disabled with
                        K8STOOLS_REDACT=0)
  --log-level {DEBUG,INFO,WARNING,ERROR,CRITICAL}
                        Log level [default: INFO]

It finishes with a summary of what it got:

$ k8s-capture-state --namespace default -o incident-1234.json
Captured 27 pod(s), 31 container(s).
  logs captured:          31
  previous logs captured: 31
  values redacted:        4
Wrote incident-1234.json (1,435,335 bytes, redacted).

What gets captured

Everything the tools can read, so that every tool answers on replay:

A cluster is selected exactly as for the MCP server (see Selecting a cluster), and a selection that cannot be loaded fails the capture.

--namespace restricts only the namespaced resources; namespaces and nodes are always captured in full so the cluster still makes sense.

Logs dominate the file size — expect a few MB for a few dozen pods (the 27-pod capture above was 1.4 MB with only 50 lines per container; a 38-pod one at 200 lines was 3.3 MB). Use --max-log-lines to trade size against history, or --no-logs for a structure-only snapshot. Captures compress about 10x, and can be written and read gzipped — see Compression, and what to check in.

Compression, and what to check in

A name ending in .gz is written gzipped, and captures are read back compressed or not either way — detection is by content, so a capture still loads after being renamed:

k8s-capture-state -o incident-1234.json.gz     # 3.3 MB -> 364 KB
k8s-mcp-server --state-file incident-1234.json.gz

For a fixture you check into git, use plain .json anyway. Git already stores blobs compressed, so you do not pay the uncompressed size — and it deltas successive versions of a text file against each other, which it cannot do for a gzip blob. Measured on two captures of the same 38-pod cluster:

The two cost about the same once, but every re-capture of a .gz adds a whole fresh blob, so the repo grows roughly 6x faster. Captures are also indented rather than compact for the same reason — it gives git something to delta. This is also why you probably do not need git-lfs here: LFS stores every version whole, which gives up exactly the delta compression that makes these files cheap.

Reach for .gz when the file travels on its own — attached to an issue, dropped in object storage, or simply too large to keep expanded in a working tree.

Previous-instance logs

The capture includes the previous-instance logs of every restarted container — what get_logs_for_pod_and_container(..., previous=True) returns. For a crash-looping container the current instance's logs are usually empty or post-restart, and the stack trace or OOM message that explains the crash lives in the previous instance, so a capture without them would be missing the evidence that matters most.

They roughly double the log payload for a cluster where a lot is restarting — which is exactly the cluster worth capturing. --no-previous-logs skips them. The summary reports how many could not be read (the previous instance may have been garbage-collected), because a crash-loop capture that lost them looks identical to one that never had any.

Previous-instance log semantics

previous=True asks the kubelet for the container's last terminated instance. Three of its behaviors read as bugs and are not, and all three show up in captures as well as in live queries:

• In CrashLoopBackOff, both calls return the same text. There is no running instance during the backoff, so the kubelet serves the most recently terminated one for previous=False too. Check get_pod_container_statuses before concluding the flag was ignored.
• A restart between two calls shifts the window. The instance that was current becomes the previous one, so a container crash-looping every few seconds can return byte-identical output for both. Compare the log timestamps, not the flag.
• A reclaimed log file returns 200, not an error, with the body unable to retrieve container logs for <container-id>. It is a successful call whose content is an error message.

Only the single most recent terminated instance is retained; there is no way to reach further back than one.

Replay clock

Every age in a capture is stored relative to the moment of capture, so the intervals between resources survive replay exactly: a deployment aged 8 days whose newest replica set is aged 7h34m still says "upgraded 7h34m ago" whenever it is replayed.

By default those ages advance in real time from when the server starts, which is what makes an interactive session feel like a live cluster. For an automated test suite pass --state-time frozen:

k8s-mcp-server --state-file incident-1234.json --state-time frozen

which pins every age at its captured value, so repeated runs are identical and a long session cannot see ages drift between its first tool call and its last. (--state-time requires --state-file; it is rejected rather than ignored, since a suite that believed it had frozen the clock and had not would be flaky with no visible cause.)

Absolute times are replayed the same way: every timestamp a tool returns is moved so the moment of capture becomes the moment the server started. A container killed 5 minutes before the capture reads as killed 5 minutes before server start, and get_cluster_info's captured_at is that start time, not the date in the file, so the agent sees one consistent clock. Log line timestamps (the kubelet's prefix) are moved too, so since_seconds works on a capture of any age; timestamps an application writes inside its own log messages are left as recorded. The file's own capture date is in the server's startup log.

Replaying from Python

For a test suite, skip the server and load the capture directly. MockState has the same query methods as the tools:

from pathlib import Path
from k8stools.mock_state import MockState

state = MockState.from_file(Path("incident-1234.json"), frozen=True)

pods = state.get_pod_summaries("default")
logs = state.get_logs_for_pod_and_container("ad-647b4947cc-s5mpm", "default",
                                            previous=True)

Or point the mock_tools module at it, so anything already calling k8stools.mock_tools serves your capture instead of the built-in one:

from k8stools import mock_tools

mock_tools.load_mock_state(Path("incident-1234.json"), frozen=True)
agent = Agent(model="openai:gpt-4.1", tools=mock_tools.TOOLS)

Captures and secrets

k8s-capture-state applies the same redaction pass as the MCP server, on by default, opt out with --no-redact or K8STOOLS_REDACT=0. This matters because a capture file is far more portable than cluster access — it gets committed to git, attached to issues, and passed around long after anyone remembers which cluster it came from. Redaction is one-way: replaying a redacted capture with --no-redact restores nothing, so a scenario that genuinely needs real values has to be re-captured.

Secret redaction

Some read-only resources can carry secret-shaped values even though they are not Kubernetes Secret objects — ConfigMap data and the env blocks in a pod spec are the common cases. When you run the k8stools MCP server directly against an agent (with no wrapping service to scrub output), those values would otherwise flow straight into the model's context.

To prevent that, the MCP server applies a redaction pass to every tool's output. It is on by default and can be disabled with --no-redact or by setting K8STOOLS_REDACT=0. Redaction matches two ways, replacing each match with a visible [REDACTED] marker so the agent can tell "hidden" from "absent". We never provide a reader for Kubernetes Secret objects.

1. By value shape — AWS access keys, JWT / bearer tokens, PEM private-key blocks.

2. By key / env-var name — the name contains key, secret, token, password, passwd, credential or cred as a whole word. Names are split on separators and camelCase, so AWS_SECRET_ACCESS_KEY, apiKey and x-auth-token all match while VALKEY_ADDR does not.

   A field named exactly key is exempt: in the Kubernetes API that is always a map entry, a taint, a label selector or a projected-volume item, never a credential. This matters more than it sounds — on a real 38-pod cluster it was 130 of 139 redactions, none of them secrets: toleration keys such as node.kubernetes.io/not-ready, the filename ca.crt, and secretKeyRef.key. That last is worth stating plainly: a reference to a secret is not a secret. The value lives in the Secret object and is resolved by the kubelet, never appearing in tool output, so redacting the pointer hides which key feeds an env var and protects nothing. The exemption does not apply to env-var names, where a variable someone named KEY plausibly does hold one.

If you call the tool functions directly in Python (rather than through the MCP server), you get raw, un-redacted values; you can apply the same pass yourself via k8stools.redaction.redact_object.