k9s taught us to live in the terminal. k10s makes that terminal point-and-click, instantly searchable, themeable - and gives it an AI that already knows your cluster, namespace and selected object.
Install · Try it with no cluster · Features · k9s → k10s · Docs · Build log
A real terminal capture of k10s running - just screenshot, on the offline demo backend. Every frame in these docs comes out of the running binary, never hand-drawn.
A cluster dashboard is something you open twenty times a day, usually while something is on fire. The two things that matter are how fast it opens and how little you have to remember.
- Nothing is hidden behind memorised keys. The actions that apply to the thing you selected are listed, right there, in their own pane. Click one, or press the letter next to it.
- Your mouse works. Click a row, click a pane, click
[ zoom ], clickns default ▾, scroll the table. Every border button is real. - One search box for the whole cluster.
ctrl+psearches resource kinds and objects together. No prefix language to learn. - It opens instantly. Startup registers zero watches and waits for nothing - informers start lazily, per kind, the first time you look at one. Opening Pods watches pods, not every Secret and Event you own. (how, and the regression guards)
- AI that can see the screen.
ctrl+a, ask in plain English. The current context, namespace, kind and selected object are injected into the prompt, so "why is this pod unhealthy?" means this pod. - It updates itself.
/updateinstalls the newest release over the running binary - checksum-verified, atomic, offers to restart into it.
k10s ships an offline demo backend. No kubeconfig, no cluster, no risk - you get a realistic cluster to click around in, including a CrashLoopBackOff to poke at.
git clone https://github.com/p10node/k10s && cd k10s
go run . # no cluster reachable → offline demo modeThe same binary connects to your real cluster the moment one is reachable.
There is no flag to remember: main.go builds the live client-go backend and
falls back to the demo only if that fails.
go install github.com/p10node/k10s@latestOr from a clone, which also stamps the version into the binary:
git clone https://github.com/p10node/k10s && cd k10s
just install # → $GOBIN/k10s, version-stampedOnce a release is tagged, prebuilt static binaries for darwin/amd64, darwin/arm64, linux/amd64, linux/arm64 and windows/amd64 are published automatically, and every later upgrade is just:
k10s updateThen run it:
k10s # your current kubeconfig context
k10s --version # which build is this16 kinds across Workloads · Network · Config · Storage · Cluster · Custom Resources, each with a live row count for the current namespace. Your CRDs are discovered automatically - no configuration.
Pods · Deployments · StatefulSets · DaemonSets · Jobs · CronJobs · Services · Ingresses · ConfigMaps · Secrets · PVCs · Nodes · Namespaces · Events · CRDs · Custom Resources
d describe |
real kubectl describe output, from the API |
y YAML |
the live object |
l logs |
follows (-f), newest at the bottom, scroll back 500 lines |
s shell |
a real interactive exec session - raw TTY, resize-aware, in-panel |
p port-forward |
real SPDY forward, start/stop from the pane |
m top |
pod/node metrics, per container, with requests vs limits |
e edit · r restart · c scale |
rollout restart, scale, $EDITOR |
o / u |
cordon-uncordon / drain - offered only when Nodes is selected |
D delete |
red confirm modal, because it should be scary |
╭─ logs -f billing-worker-6f8d9c5b7-qq91x ────────────────────────────────── [ close ] [ restore ] ╮
│ 8 2026-08-25T08:12:18.331Z INFO http GET /v1/users/me 200 3.1ms │
│ trace=44b1e2f9 │
│ 7 2026-08-25T08:12:21.660Z INFO worker flushed batch size=250 dur=41ms │
│ 6 2026-08-25T08:12:25.019Z INFO http GET /healthz 200 0.3ms │
│ 5 2026-08-25T08:12:31.402Z INFO http DELETE /v1/sessions/9a1 204 5.7ms │
│ trace=7c0d19ba │
│ 4 2026-08-25T08:12:33.881Z WARN gc pause=18ms heap=412Mi │
│ 3 2026-08-25T08:12:40.117Z INFO http GET /v1/orders/88213 200 7.4ms │
│ trace=e21f8b05 │
│ 2 2026-08-25T08:12:44.590Z INFO metrics scrape ok series=1842 │
│ 1 2026-08-25T08:12:51.008Z INFO http GET /healthz 200 0.3ms │
│ ● following newest at bottom 500 loaded · ↑ for older │
╰──────────────────────────────────────────────────────────────────────────────────────────────────╯
╭─ Prompt · plain text → AI · /commands still work · esc close ─ [ grow ] [ AI · claude-sonnet-5 ] ╮
│ ✦ ask about your cluster… · /settings to change provider/model │
╰──────────────────────────────────────────────────────────────────────────────────────────────────╯
● AI mode — plain text goes to claude-sonnet-5 tab panes · enter open · ctrl+p search · f find…
ctrl+a toggles it. Bring your own key - OpenAI-compatible (so also
Groq, Together, OpenRouter, vLLM, Ollama, LM Studio, any local gateway) or
Anthropic. Answers open as a normal text view you can scroll, zoom and
close. No cluster data leaves your machine unless you press enter in AI
mode. The only other network call k10s makes on its own is the once-a-day
update check, which asks GitHub for a version number and nothing else.
tokyo-night (default) · catppuccin-mocha · dracula · nord ·
gruvbox-dark · solarized-light · matrix
T cycles, /theme opens a picker that applies each theme as you move
through the list - you judge it on the real UI, not on a name - and esc
puts back whatever you had. No hardcoded color exists anywhere in the render
path, so adding a theme is one struct.
- Copy mode (
ctrl+s) releases the mouse so your terminal can drag-select and copy - the one thing every mouse-capturing TUI breaks. - An honest loading state instead of "no resources found" while a watch is still syncing.
- Namespace and context pickers that never ask you to type a name.
- First-run onboarding asks what you call kubectl (
kubectl/k8s/k/ your own alias) and uses it in every hint from then on. - A grow-able prompt (
ctrl+z) - because a long kubectl line or an AI question does not fit in a one-row field that scrolls sideways. - A terminal-too-small notice instead of a garbled layout.
k10s exists because of k9s. Credit where
it is due: k9s is mature, enormous in scope, plugin-extensible, and it is the
reason a whole generation of us stopped typing kubectl get pods all day.
k10s is younger and deliberately narrower. The difference is philosophy:
| k9s | k10s | |
|---|---|---|
| Primary input | keyboard-first, command-driven | mouse and keyboard, equally |
| Discovery | you learn the commands | the actions for the selected object are listed on screen |
| Search | per-view filter, plus : to jump between views |
one box (ctrl+p) over kinds and objects |
| AI | plugin territory | built in, with cluster context injected |
| Scope | very large, plugin ecosystem | small, opinionated, no plugins |
| Upgrades | your package manager | k10s update, self-replacing |
Use k9s if you want the biggest feature surface, a plugin ecosystem, and
years of production mileage. Use k10s if you want something you can hand
to a teammate who has never opened a TUI, and have them find logs on their
own in ten seconds.
The short version - the full reference is here.
| Key | Action |
|---|---|
tab / shift+tab |
cycle panes |
↑↓ (j/k) · pgup/pgdn · g/G |
move · page · first/last |
| type (Resources focused) | instant kind filter · esc clears |
f |
find: search every column of the open table |
ctrl+p |
search everything - kinds and objects in one box |
enter / double-click |
open it: logs if it has them, else describe |
z |
zoom / restore the centre pane |
: / / |
command prompt (with a suggestions popup) |
ctrl+a |
AI prompt mode ✦ |
T / ctrl+t |
cycle themes |
ctrl+s |
copy mode - release the mouse to select & copy |
| click / wheel | anywhere in a pane selects it, blank space included |
q / ctrl+c |
quit |
Two prefixes, and each one only ever shows its own set. / does
something. : narrows what is on screen. enter runs the highlighted
suggestion immediately - no second trip through the prompt.
/ns /context /theme /settings /update /version /help
:search <term> :filter <term> :scale <n> :mouse
None of /ns, /context or /theme takes a name - each opens a chooser
showing what actually exists. Full command reference.
Theme, context, namespace, AI settings and the update check persist to
~/.k10s/config.yaml (override with K10S_CONFIG), saved on every change.
See config.md.
Note: the AI API key is stored as plain text in that file (mode 0600) and masked in the UI. If that is not acceptable in your environment, leave it empty and AI mode simply stays off.
| architecture.md | packages, the Source boundary, render pipeline, mouse zones |
| performance.md | lazy watches, cheap counts, and the guards that keep it fast |
| ui.md | layout, panes, focus, zoom, overlays |
| keybindings.md | full key + mouse reference |
| commands.md | prompt modes, slash commands, AI settings |
| themes.md | theme list, picker, palette contract |
| config.md · update.md | what persists · how self-update works |
| backends.md · dev.md | live vs demo backend · build, tests, headless renderer |
| roadmap.md | what is done, what is missing, and the known limits |
| marketing.md · build-in-public.md | how this project is being shared |
The barrier is low on purpose. just lists every recipe:
just dev # run from source
just shot 140 44 # render one frame headlessly - no TTY, no cluster
just screenshot # capture the README hero from a real terminal (vhs)
just check # fmt + vet + test, in the order that fails fastestcmd/shot is why UI work here is pleasant: you can iterate on layout, replay
keystrokes (just shot 140 44 j,j,d) and diff frames without a terminal or a
cluster. Nothing in these docs is drawn by hand - the frames come from that
renderer, the hero image from just screenshot driving the real binary in a
real terminal.
Good first issues live in roadmap.md → possible next steps: more kinds, click-to-sort columns, multi-select bulk delete, grep inside the log stream, saved views.
Honest status: every live path is implemented against the real
client-go/kubectl APIs and covered by tests with fake clientsets, but this
tree has not yet been smoke-tested against a production cluster by its
author. Try it on kind first. The known limits are listed
in the roadmap rather than buried.
To my colleagues - the people whose questions, over-the-shoulder debugging sessions and "can you just check if that pod is up?" pings are the entire reason this exists. k10s was not built to be a project; it was built so that looking at our clusters would stop being the annoying part of your day. Every pane in it is an answer to something one of you asked me. Thank you for the inspiration, the patience, and for being the first users.
Cảm ơn các đồng nghiệp của tôi - repo này được viết ra để phục vụ các bạn, và chính các bạn là nguồn cảm hứng cho k10s.
To @derailed and k9s, for proving a Kubernetes TUI could be something you actually want to open.
To Charm - Bubble Tea, Lip Gloss, Bubbles - and to BubbleZone, which is the reason any of this is clickable at all.
Apache-2.0 - the same license as Kubernetes, Helm and k9s. Use it, fork it, ship it inside your company.