lfk

lfk (Lightning Fast Kubernetes navigator) is a Bubbletea-based TUI for Kubernetes that swaps the single-list k9s model for a yazi-inspired three-column Miller view. The left column is the parent level, the middle is the current selection, the right is a YAML or details preview — navigation walks down a single hierarchy: Cluster → Resource Type (grouped) → Resource → Owned Resources → Containers. Namespaces are not a navigation level; they live in a top-right indicator and switch via \.

Listed on watchlist — six-week-old repo (created 2026-03-15), single author (janosmiko), pre-1.0 (v0.9.30 at time of ingest), 320 stars, but pushing 30 releases in six weeks and active CI with SonarCloud + codecov. The shape is mature; the version number is not.

What's actually different from k9s

The flagship feature is owner-based navigation. Selecting a Deployment shows its Pods (resolved through ReplicaSets, flattened); a CronJob shows its Jobs; a Service shows Pods matching its selector; an ArgoCD Application shows its managed resources (status- or label-derived); a Helm release shows resources via the app.kubernetes.io/instance label. You don't list "all pods then filter by owner" — you walk into the owner. o jumps the other way, from a resource to its controller.

Resource types are grouped semantically: Workloads, Networking, Config, Storage, ArgoCD, Helm, Access Control, Cluster, Custom Resources, plus Dashboards. CRDs are auto-grouped by API group (argoproj.io, longhorn.io, networking.istio.io); pin frequently used groups with p (per-context). Rarely used types (CSI internals, admission webhooks, APF, leases, runtime classes) are hidden by default and surface under "Advanced" with H.

What's built in vs what shells out

Only kubectl is a hard dependency. Optional: helm for release management, trivy for image vulnerability scans. Everything else — KEDA pause/unpause, External Secrets force-refresh, Argo Workflows submit/suspend/stop/resubmit, ArgoCD sync/terminate/refresh, cert-manager, FluxCD, PVC resize — talks to the Kubernetes API directly. No per-tool CLI.

Embedded PTY for exec/shell stays running when you switch tabs (this is rare in TUI tools — most kill the PTY on tab switch). Inline log viewer streams across init-container transitions, staying attached as each finishes and the next starts.

The Secret problem and lazy loading

In a cluster with many Helm releases or large TLS secrets, listing the Secrets resource type can transfer tens of megabytes because k8s returns full bodies in a list call. secret_lazy_loading: true in config switches to metadata-only listing, deferring decoded values to hover (cached after first fetch). Trade-off: the Type column drops from the list because the metadata-only fetch doesn't populate it. This is a real operational concern — the kind of thing you only hit once you've used the tool against a real cluster.

Theming surface

460+ ghostty color schemes ship in (Tokyonight, Catppuccin, Dracula, Nord, Rose Pine, Gruvbox, …). T opens a live picker. Auto OS dark/light switching uses CSI 996/2031, supported in Ghostty, kitty ≥0.27, Contour — colorscheme: "dark:catppuccin-mocha,light:catppuccin-latte" and the app flips when the system appearance changes. Transparent background is a checkbox. Compare against k9s's "edit a YAML file and restart."

Icon mode is auto by default with explicit fallback to unicode / nerdfont (Material Design Icons) / simple (ASCII labels) / emoji / none. Override with LFK_ICONS env var.

Command bar surface

: opens a command bar with vertical-dropdown autocomplete. Four overlapping syntaxes:

• Resource jumps: :pod, :dep, :pvc — fuzzy-matchable abbreviations, configurable
• Built-ins: :ns, :ctx, :set, :sort, :export
• kubectl passthrough: :k get pod or :kubectl ... with flag/namespace completion
• Shell: :! cmd

Search and filter share three modes auto-detected from the input: substring, regex (auto-detected), fuzzy (~ prefix), literal (\ prefix). Multi-select with Space, range-select with Ctrl+Space, bulk delete/scale/restart and ArgoCD bulk sync/refresh from the action menu.

Configuration

~/.config/lfk/config.yaml. Override via -c /path/to/config.yaml. CLI flags: --context, -n (repeatable for multi-namespace), --kubeconfig, --no-mouse, --no-color (also respects NO_COLOR=1), --watch-interval (Go durations, clamped to [500ms, 10m]).

Merged kubeconfig loading: reads ~/.kube/config, every file in ~/.kube/config.d/* recursively (symlinks followed), and the KUBECONFIG env var. Last-context/namespace/resource is persisted across restarts; explicit --context / -n flags ignore the saved session.

Custom user-defined actions are first-class — define shell commands per resource type in config and they show up in the action menu (x).

Install

brew install janosmiko/tap/lfk           # macOS / Linux Homebrew
nix profile install github:janosmiko/lfk # Nix flake (requires flakes)
go install github.com/janosmiko/lfk@latest

A nightly tap (janosmiko/tap/lfk-nightly) and janosmiko/lfk:nightly[-YYYYMMDD] Docker tags track main. Docker usage needs -v ~/.kube:/home/lfk/.kube:ro and --net=host for port-forwards to work.

Status and re-check signals (next: 2026-07-26)

• Created: 2026-03-15 (~6 weeks old at ingest)
• Stars: 320
• Releases: 30 in six weeks (last v0.9.30, 2026-04-27)
• Open issues: 3
• Single author, Apache-2.0
• README says nothing about being alpha; SonarCloud security rating + codecov + Go Report Card all wired into CI

Re-check signals at 2026-07-26: did v1.0 ship? Did the release cadence stay ≥monthly? Any second meaningful contributor besides janosmiko? Did anyone write a serious comparison vs k9s? Did secret lazy loading and ArgoCD bulk operations attract real-world bug reports (a sign people are using it on big clusters)? Has the test/lint surface stayed green in CI or rotted?

Repo: https://github.com/janosmiko/lfk — 320 stars, Apache-2.0.