A terminal UI for Talos Linux Kubernetes clusters: nodes, services, logs, etcd, network, storage and diagnostics in one place, with safe node operations.
talos9s is a fork of talos-pilot by Ken Udovic, continued with storage views (LINSTOR/DRBD, ZFS), views across a cluster's nodes and more. It is a community project, not affiliated with Sidero Labs.
Talos has no SSH; everything goes through its API, and talosctl is a long
list of subcommands. talos9s puts that API, and the Kubernetes state that
depends on it, on one screen. It complements k9s rather
than replacing it:
| Layer | Answers | |
|---|---|---|
| k9s | Kubernetes (:6443) |
"Why won't my pod start?" |
| talos9s | Talos OS (:50000) and the cluster around it |
"Why won't my node join?", "Is it safe to reboot this node?" |
Checked against each project's source and README in November 2026 (talos-pilot 0.1.11, its last release; t9s 0.2.0, its Homebrew cask).
| talos9s | k9s | talos-pilot | t9s | |
|---|---|---|---|---|
| Layer | Talos OS and the cluster around it | Kubernetes | Talos OS | Talos OS |
| State | Released, active | Mature, very active | Released, upstream of talos9s | Young (22 stars, no tagged release) |
| Language, UI library | Rust, ratatui | Go, tview | Rust, ratatui | Go, Bubble Tea; wraps talosctl |
| License | MIT | Apache-2.0 | MIT | Source-available, non-commercial |
| Multi-cluster, all nodes at once | yes | one context at a time | yes | one context, switched with x |
| Views across a node group (logs, network, disks, diagnostics) | yes | – | node by node | – |
| Logs, processes, network, etcd | yes | pod logs | yes | logs, dmesg, processes, containers, addresses; no etcd |
| Packet capture | yes | – | yes | – |
| KubeSpan | yes | – | peers in Network | – |
| Disks and Talos partitions | yes | – | disks | disks |
| LINSTOR/DRBD and ZFS LocalPV checks | yes | – | – | – |
| Machine config viewer (secrets masked) | yes | – | – | yes |
| Diagnostics with suggested fixes | yes, with the fix command for you to run | Popeye sanitizer | yes | talosctl health, streamed |
| Drain, reboot, shutdown, rolling | yes | cordon, drain, delete, edit, shell, port-forward | drain, reboot, rolling | reboot, shutdown |
| Talos and Kubernetes upgrades | – (planned) | – | – | yes, with the extension catalog |
| Operation confirmation and audit log | yes | confirmation; --readonly flag |
yes | confirmation for upgrades |
| Mouse | clicks, double-clicks, wheel | optional, off by default | – | – |
| Remembers where you left off | yes | per-context config | – | – |
| Skins, plugins, custom hotkeys | – | yes | – | – |
| Needs a kubeconfig | no, fetched from a node | yes | no | no |
k9s answers questions about workloads, so use it next to the Talos tools, not instead of them. Compared with talos-pilot, talos9s adds the storage checks, the group and cluster views, the machine config viewer, the mouse and the remembered state, and tightens safety (suggested fixes are never applied, reboots check etcd members that are already down).
Prebuilt binaries (macOS, Linux):
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/olexiyb/talos9s/releases/latest/download/talos9s-installer.sh | shWindows (PowerShell):
powershell -ExecutionPolicy Bypass -c "irm https://github.com/olexiyb/talos9s/releases/latest/download/talos9s-installer.ps1 | iex"Or download an archive from Releases.
From source (needs Rust 1.95 or newer; rustup update gets it):
cargo install --git https://github.com/olexiyb/talos9s --locked talos9sWith Nix:
nix run github:olexiyb/talos9s # try it
nix shell github:olexiyb/talos9s # put it on PATHYou need a talosconfig (~/.talos/config or $TALOSCONFIG) and network
access to your nodes on port 50000. The Kubernetes views fetch a kubeconfig
through the Talos API, so no separate kubeconfig is needed. A context in your
KUBECONFIG with the same cluster CA is used instead when it works.
To see and act on Kubernetes as your own identity (an OIDC user, a
read-only role) instead of the admin one, pass --kubeconfig (or set
TALOS9S_KUBECONFIG). Only that file is used, with the context whose CA
matches the cluster; when it has none, or its user may not list something,
the views say so rather than fall back to admin.
talos9s # every context in the talosconfig
talos9s --context homelab # one context
talos9s --tail 1000 # lines of log history to load
talos9s --kubeconfig ~/.kube/oidc.yaml # Kubernetes as that kubeconfig's user
talos9s --debug --log-file ~/talos9s.logWithout --log-file, each run logs to <temp dir>/talos9s-<pid>.log; logs
older than a week are removed at startup.
New cluster on machines in maintenance mode? The bootstrap wizard generates the configs, applies them and bootstraps etcd:
talos9s --insecure --endpoint <node-ip>The cluster tree and the menu stay on the left; the chosen view fills the right side for whatever is selected. Move the selection and the view follows (a node's network, then the next node's, or the whole group's on a group row). Recently shown views come back instantly.
The menu groups the views: Overview (dashboard, diagnostics, logs, processes), Node, Network, Storage and Cluster. The Node entries (containers, machine config, extensions) need a node and are dimmed on a cluster or group row.
The Dashboard shows:
- All contexts at once, each with its control plane and workers, node health and etcd quorum.
- Per cluster: a node table (versions, memory, disk, ZFS, load, uptime, services, clock offset, gateway), the KubeSpan peer matrix and LINSTOR nodes and pools.
- Per node: resources, addresses, routes, DNS, Tailscale, KubeSpan peers, clock, and every Talos service with its state.
- Service logs, searchable and colored by level; several services interleaved (Stern-style). JSON and logfmt lines (kubelet, zap, slog, Pino ...) read as their message, then the other fields.
- On a cluster, control plane or workers row: the same services from every node merged by time with a node column, including the kernel log.
| View | Shows |
|---|---|
| Processes | htop-like process list and tree, sorted by CPU or memory |
| Network | Interface traffic and connections, packet capture (filtered so it doesn't capture itself) |
| KubeSpan | The node's WireGuard peers; on a cluster or group row, the peer matrix |
| Disks | The node's disks, read-only ones marked RO; virtual devices (loop, dm, md, DRBD, zvol) hidden until A, which is remembered |
| Partitions | Talos volumes: the partitions with their filesystem, encryption and mount point |
| Containers | Every container on a node, a group or the cluster, read from containerd through the Talos API, so it works when the Kubernetes API doesn't: one row per container, Talos services included, exited ones hidden until A, sort by any column (1-7 or a header click), W namespace warp. Enter shows one container with its logs, as in Workloads (levels, search, wrap, V to select and copy lines, v opens them in hl); y copies the talosctl logs command |
| Machine config | The config in effect, read-only, searchable, secrets masked; flags a staged change waiting for a reboot; the outline shows each part's size in bytes, # switches to lines |
| Extensions | Installed system extensions |
| Diagnostics | Health checks with suggested fixes, etcd member checks on control plane nodes (apply lag, behind the leader, fragmentation with the defrag command, quota), CNI checks (Flannel, Cilium, Calico) and addon checks (cert-manager and more) |
On a cluster, control plane or workers row, network, storage and diagnostics show every node at once; diagnostics as a node × category grid with each warning listed once with its nodes.
| View | Shows |
|---|---|
| etcd | Quorum, members, leader, alarms, advertised and listen addresses (running vs config), live member log |
| Volumes | LINSTOR/DRBD and ZFS LocalPV volumes, with problems first (see below); W warps into a PVC namespace, shared with Workloads |
| Workloads | Unhealthy Deployments, StatefulSets, DaemonSets and pods in one list, worst first; NODES shows where each runs, or for a failing one, where its unready pods are; A shows healthy ones too, S groups by namespace, / filters; W warps Workloads and Volumes into namespaces (a picker: Space marks several, prefix* warps by prefix, the last 4 warps kept per cluster, 1-4 pick them, 0 all); P lists every pod instead (ready, node, state, restarts, age, owner; static and Job pods too), a click on any column header or 1-8 (the columns left to right) sort it, by state (worst first) by default; below the list, the selected pod's containers, last termination and latest event, and y copies the kubectl command. Enter focuses a workload: its pods, and the logs of all of them merged by time; select a pod for its containers' logs, Enter on it lists its containers, init containers too, select one for only its logs (p the previous run, where a container restarted, w wrap, 1-5/0 levels as in Logs, / search with n/N, PgUp/PgDn/g/G or the wheel scroll, follows the end; w wraps long lines under the message; V selects lines, j/k extend, y copies them as shown, while a bare y copies the kubectl logs command; v opens the lines read in hl, or the viewer TALOS9S_LOG_VIEWER names), levels coloured, JSON and logfmt lines shown as message and fields, terminal colour codes removed, the date shown when the lines span days |
| Kubernetes | Every cluster of the talosconfig in one list, read-only: the unhealthy Deployments, StatefulSets, DaemonSets and pods of all of them (a CLUSTER column; A shows everything), any kind from API discovery, CRDs included (: picks one: the common kinds until you type, then every kind that matches, with the clusters serving it), or a search (s: a name, l:app=web a label selector, i:redis an image). Each cluster is read on its own: one that can't be reached or answers 403 shows ? with its error, the rest are listed. X shows one cluster at a time, / filters. Enter shows an object: status, conditions, events and its YAML, masked (Secret values never leave the API answer). d diffs it with the same object in another cluster (or marks it, and d on a second row compares the two). l reads its pods' logs, L those of the same selector in every cluster, or of every pod a label search found. y copies the kubectl get -o yaml command |
| GitOps | Flux Kustomizations, HelmReleases and sources and Argo CD Applications across clusters, worst first: health and sync (out of sync counts as degraded, suspended is called out), the applied revision and the source (repo, path or chart, target revision) |
| Lifecycle | Talos and Kubernetes versions across nodes; a saved config that differs from the one in effect is flagged "staged, reboot" |
| Security | Certificate expiry, encryption status |
Volumes compares LINSTOR's view with each node's kernel and flags what LINSTOR alone doesn't show:
- DRBD devices or filesystems not grown after a resize;
- out-of-sync blocks on connected links, including those that can make a Primary fail its I/O after a short link loss (LINBIT/drbd#153);
- disconnected and StandAlone links (split-brain), suspended I/O, missing replicas, resync progress with an ETA, filesystems filling up;
- orphaned volumes: no PersistentVolume uses them, or their claim is gone.
Where a fix is provably safe, it suggests the exact commands to copy (y).
talos9s never runs them.
The Operations view (o) lists what can run on the selected node, group or
cluster, and every run so far. On a group or cluster an operation runs node by
node and stops at the first failure.
| Operation | What it does |
|---|---|
| Cordon / Uncordon | Stop new pods on the node(s), keeping the running ones / allow them again |
| Drain | Cordon and evict pods, respecting PodDisruptionBudgets; stays cordoned |
| Reboot | Drain, reboot, wait for Ready with a new boot ID, uncordon; on a group, one node at a time with etcd quorum checked before each control plane node |
| Shutdown | Drain, power off; stays cordoned until you uncordon it (one node only) |
- Starting: pick the operation:
o(orEnteror a click on it) opens the drop list,1-5pick one directly; the cursor then moves to the nodes. Check them withSpaceor a click (on a group none is checked, "all nodes" checks every one; a single node is checked already).Enteropens the confirmation: the checked nodes, the equivalentkubectl/talosctlcommands, etcd quorum and blocking PDBs;ystarts it,n/Esccancels. The history's TARGET is the group when all its nodes ran, else the nodes. - Node state: each node row shows TALOS (up and its uptime, or
offline), K8S (Ready), SCHEDULE (cordoned or not), PODS (what a drain
evicts) and its last run, read every 15 s and when a run ends. A node the
operation can't run on shows
[-]and why: Cordon needs an uncordoned node, Uncordon a cordoned one, Reboot and Shutdown a node answering the Talos API (Cordon and Drain also work on a node that is down). A state that can't be read blocks nothing. - Drain options (Drain, Reboot, Shutdown): the confirmation fills the
screen, tree included (a click elsewhere can't lose it), with the options
on top and a table of every pod on the checked nodes and what happens to
it (stops the drain, blocked by a one-instance CloudNativePG cluster,
waits for a PodDisruptionBudget that allows no disruption now, gone for
good, emptyDir data lost, a CloudNativePG primary that switches over
first, evicted, stays), worst
first;
PgUp/PgDnand the mouse wheel scroll it, and the pods the option under the cursor decides about are bold.rreads the pods again, and they are read every 10 s (atoggles). Ignore DaemonSets, Delete emptyDir data and Force check boxes, grace period, timeout (5 min per node), Wait to finish (off, 5 min to 1 h: after cordoning, wait for pods no controller recreates, such as CI agents, and with "also Job pods" for Jobs, to end by themselves before draining; what is left then follows Force) and, for a reboot of several nodes, the wait between them.j/kmove,Spacetoggles,←/→change a value, a click toggles. Each counts what it hits on the checked nodes (15 pods stay,stops the drain: 1 pod). As with kubectl, a pod an option doesn't allow (a DaemonSet pod, one no controller recreates, one with emptyDir data) stops the drain before the node is cordoned. The choice is kept per cluster, except Force, which starts off every time. - A node cordoned before a Drain, Reboot or Shutdown stays cordoned: a run uncordons only what it cordoned itself.
- Runs in the background: switch views or nodes freely; the header shows
the run and its step on every view. One run at a time per cluster context.
xcancels after the current step (a failed or cancelled node is uncordoned again). While a run goes,qandCtrl+Cdon't quit; a secondCtrl+Cwithin 2 seconds does. - History: every run is kept in
~/.talos9s/operations/(one private file each: time, context, nodes, commands, every step, how it ended) until you delete it withD. On a node the list shows that node's runs (A: every run); on a group or cluster every context's (A: this one only),/filters by node, operation, status or command,Entershows a run's output (search, wrap,V/ylines,vviewer),ycopies its commands.
Every step also goes to the audit log.
In the tree (each key switches the view on the right; the left menu and the mouse do the same):
| Key | On a node | On a cluster, control plane or workers row |
|---|---|---|
D |
Dashboard: node details | Dashboard: cluster or group summary |
l / L |
Logs | Logs merged across the nodes |
n |
Network | Network of every node |
K |
KubeSpan peers | KubeSpan peer matrix |
s / t |
Disks / partitions | Disks / partitions of every node |
d |
Diagnostics | Diagnostics grid |
p |
Processes | Processes |
C |
Containers | Containers of every node |
m / x |
Machine config / extensions | |
o |
Operations: cordon, uncordon, drain, reboot, shutdown; run history | Same, node by node (no shutdown) |
v / w |
Volumes / workloads on this node | Volumes / workloads on those nodes |
i / g |
Kubernetes / GitOps across every cluster | same |
e b c |
etcd, lifecycle (boot and upgrades), security (cluster-wide) | same |
Everywhere:
?lists every key that works now: the view's own (as on the bottom line), the view letters, the tree's and the shared ones.j/kscroll it,Esc,qor?close it.j/kor arrows move the selection and the view follows;g/GorHome/Endjump to top or bottom,PageUp/PageDownmove a page.→,Enteror a click puts the keyboard in the view;←,Escorqgives it back (qdoubles forEscbecause some terminals, IntelliJ's among them, never passEsc).Tab/Shift+Tabcycle Nodes → menu → view.qin a view returns to the Dashboard; there it asks before quitting (y).Ctrl+Cquits at once, unless an operation runs (then twice)./filter or search,rrefresh,aauto-refresh,Ashow all (or issues only),0all levels, services or nodes in logs,ycopy,Spacefold. The same letter means the same thing in every view; uppercase in the tree (L) is the across-nodes version of a view.[/]narrow or widen the left pane,Zzooms the view to full width.- Mouse: click a row to select it, double-click to open it, click a menu entry or a sortable column header (Network) to switch or sort; the wheel scrolls the pane under the pointer.
- The top line lists the keys that work everywhere; the bottom line only what acts in the focused window right now.
talos9s starts where you left it (see Safety for the state file).
talos9s runs with your talosconfig, which usually means full admin on every node, so it is careful about what it changes:
- Reads by default. It changes a node only for cordon, uncordon, drain,
reboot, shutdown, cluster install and bootstrap, each after
a confirmation dialog, and logs every step to
~/.talos9s/audit.log(operations also to their run file in~/.talos9s/operations/). - Never runs suggested commands. Diagnostic fixes (talosctl) and storage fixes (drbdadm, kubectl) are shown and copied for you to review and run.
- Never shows secrets. The machine config viewer masks keys, tokens, passwords, private keys and Secret data before anything is drawn.
- Writes only: the audit log, the operation runs, the remembered view state
(
~/.talos9s/state.yaml: selection, view, pane width, folded groups, sort orders and toggles, no cluster data, no text filters; delete it to start fresh), its debug log (temp dir or--log-file), packet captures you save (/tmp, owner-only) and the wizard's output directory. It has no configuration file. - No in-cluster access needed. Nothing runs in the cluster and no Kubernetes namespace needs Talos API rights.
- One connection besides your clusters: at start (at most once an hour)
and once a day while running, it asks GitHub for the latest talos9s
release, to show "0.4.0 available" on the top line.
--no-update-check,TALOS9S_NO_UPDATE_CHECK=1orDO_NOT_TRACK=1turns it off.
- State over logs. Health comes from current state (files, Talos and Kubernetes APIs), not from log lines, so an old error in a log doesn't raise an alarm.
- Unknown over wrong. If a source can't be read, a check says unknown instead of guessing; nothing crashes, the rest of the screen keeps working.
- Nodes, not the VIP. The Talos API is addressed at each node; the
cluster VIP needs a healthy control plane, so it fails exactly when you
need diagnostics. Talos resources (COSI) aren't exposed on
:50000, so they are read throughtalosctl get. Every call has a timeout, so a dead node never freezes the UI.
cargo test --all
cargo clippy --all --all-targets -- -D warnings
TALOS9S_CONTEXT=my-cluster cargo dev # rebuild and restart on every change (needs cargo-watch)The workspace has three crates: talos-rs (Talos client), talos9s-core
(logic, unit-tested without a terminal or cluster) and talos9s-tui (the
ratatui app). Git hooks, cargo deny, regenerating the Talos API client and
the other conventions are in CLAUDE.md.
- Local Talos clusters for testing: test-clusters/
- Design plans, e.g. rolling upgrades: plans/
- Rules per area: .claude/rules/
Contributions are welcome. Keep changes in line with How it works: check real state, degrade gracefully, and never report failure when "unknown" is the honest answer.
MIT, see LICENSE.
- talos-pilot by Ken Udovic, which talos9s is forked from
- Talos Linux by Sidero Labs
- k9s for the inspiration
- ratatui