A fast, native Kubernetes IDE for your desktop: your workloads, config, network, storage and custom resources in one window, a UI that reads from a live cache, and AI agents that work with you, not behind your back.
Cogpit gives every standard Kubernetes object, every custom resource your cluster serves, Helm, Flux CD and Argo CD a navigator, searchable lists, detail drawers, actions and shortcuts, in an open-source app written in Go. Every list reads from a cache that the cluster's watches keep current, so lists of 10,000 pods stay quick and opening a view again costs your API server nothing. And every action is also a tool an agent can use over MCP, with you deciding what it may change.
Alpha. Cogpit runs on macOS on Apple silicon, on Debian and Ubuntu (amd64 and arm64) and on Windows 10 and 11 (x64). Every core workflow has been walked end to end on real clusters; some views and actions are still partial and rough edges remain, so try it on a test cluster first. Where it stands, feature by feature: ROADMAP.md.
Workloads, config, network, storage, access control and every custom resource your cluster serves, each in a list you can search (plain, match case or a regular expression) and narrow to the states you care about: Running, Pending, CrashLoopBackOff, several at once. Pods and nodes show their CPU and memory from the Metrics Server or Prometheus; a pod's drawer charts them per container against requests and limits, with Prometheus up to 30 days back (pictured above).
The Overview charts CPU, memory and pods of your worker or control-plane nodes over the last hour or longer, against requests, limits and capacity, and lists what is in trouble right below: failing pods, unready nodes and workloads, warning events.
Usage bars for CPU, memory and disk (from node-exporter) in the Nodes list; a node's drawer shows its capacity and the pods running on it with their CPU and memory. Cordon it, drain it (a drain can be finished even if Cogpit quits halfway) or open a shell on it; the drain and the shell ask first.
Follow one container, or every pod of a Deployment, Stateful Set or Job in a single tab, each line marked with its pod and container. Search highlights or filters; timestamps, the previous container, older lines and a jump to a point in time are one click away.
Open a shell in any container, attach to it, or open a terminal bound to the cluster's context.
Cogpit checks your kubectl against the cluster's version and offers a matching download,
verified against the Kubernetes release's signature, when yours is missing or too far off.
Edit any object in place. Before anything reaches the cluster you see the diff and the API server's dry-run verdict, and if someone changed the object in the meantime, Cogpit tells you instead of overwriting their change. Unsaved edits survive a crash.
Browse the charts of the repositories you already added with helm repo add, with their icons,
versions and each version's README, and install from a tab. Your releases show their history,
values and notes; upgrade to a newer chart version, roll back or uninstall, all through the Helm
SDK.
Flux CD gets its own section: a dashboard of every Flux kind's state and what is not ready, with Reconcile (with its source), Reset, Suspend and Resume on one object or a whole selection. Argo CD shows your Applications' sync and health, the recent activity and what needs attention, with Sync, Refresh and a way into the Argo CD UI.
Cogpit reads every kubeconfig in ~/.kube and the files and folders you add, each on its own, and
follows their changes. Pin clusters to the hotbar (F1–F12), keep tabs of several clusters open
(they come back after a restart), and reach any view, action or cluster from the command palette
(⌘⇧P).
A Secret's drawer decodes every certificate in it, PEM or DER: subject, issuer, names, validity
with a warning 30 days before it ends, key usage and fingerprints. Cogpit also checks that
tls.key belongs to tls.crt and that the chain is in order, without ever showing the key.
Every command in Cogpit is also an MCP tool, so an agent such as Claude Code sees what you see, reads lists and logs, and navigates the window for you. Reading is free; a change waits for your OK in Cogpit's window, which names the objects, the namespace and the cluster (unless you allowed changes on that cluster; destructive ones always wait), and Help › Audit Log shows everything agents did. A few commands, such as creating a Secret, only you can run. The server agents connect to is off until you switch it on, with the MCP item in the status bar: nothing listens for agents unless you chose it.
Cogpit has an agent of its own: a chat column at the window's right edge (⌘I) that talks to a model you choose, LM Studio on your Mac, a LiteLLM proxy or Open WebUI. Ask about the pod you selected and it reads the object and its logs and answers in Markdown, tables included; each of its tool calls is a row you can open, a change waits for your OK as an MCP agent's does, and what a turn sent to the model is one click away. Escape or ⌘. stops it at any time.
A calm navy stage where only what needs you lights up, dark or light, or following your Mac's appearance, switched live with the sun and moon in the title bar or in Preferences; every time in your own time zone and your region's format. The pictures on this page follow your light or dark theme too.
- Never surprises a cluster. Browsing never writes, Cogpit installs nothing into your clusters unless you ask it to, and it never writes your kubeconfig files.
- Asks before it destroys. Delete, drain, evict, scale to zero, Helm rollback and uninstall and the node shell always ask first, naming the objects, the namespace and the cluster.
- Agents ask too. Agents read and navigate freely; anything they would change waits for your acceptance unless you allowed changes on that cluster, and destructive actions always wait.
- Everything agents do is logged, readable in Help › Audit Log.
- Secrets stay secret. Values stay masked until you reveal them, agents never read them (where a log prints a value of a Secret Cogpit holds, an agent reads a placeholder) unless you mark the built-in agent's connection private (a local model or a private cloud), and tokens and keys never reach a log.
- Your credentials stay yours. Cogpit authenticates the way
kubectldoes, with your kubeconfig, and runs a credential plugin only after you allowed it. - Light on the API server. Views read from a watch-fed cache: opening a view again costs no request.
Download the latest release:
Cogpit-<version>-macos-arm64.dmg, Cogpit for Apple silicon, signed and notarized by Apple. Open it
and drag Cogpit into Applications.
On Linux (Debian 13, Ubuntu 24.04 or later; from the release after v0.32.0 on), download
cogpit_<version>_amd64.deb (or _arm64.deb) from the same release and install it:
sudo apt install ./cogpit_<version>_amd64.debThe package also sets up Cogpit's signed APT repository, so apt and the Software Updater offer
every new release, and Cogpit's own Update button installs it after your password. Whether updates
install by themselves is your system's setting, never Cogpit's
(installing on Linux).
On Windows (Windows 10 1809 or 11, x64; from the first release after the branch
windows-release is merged), download Cogpit-<version>-windows-x64-setup.exe from the same
release and run it: it installs for you alone, without an administrator (/allusers for every
user, /S silently). It is not signed with Authenticode yet, so SmartScreen warns the first time:
More info, then Run anyway. Cogpit then updates itself: a click on its Update button downloads the
next release, checks its signature and installs it
(installing on Windows).
cogpit-cli comes inside the app, at /Applications/Cogpit.app/Contents/MacOS/cogpit-cli
(in the Linux package at /usr/bin/cogpit-cli, on Windows as cogpit-cli.exe beside cogpit.exe): it lists your contexts, prints any kind's objects
as Cogpit's views show them, and connects agents to Cogpit's MCP tools. Once you switched the MCP
server on, Help › Copy MCP command gives you the one command that registers it with Claude Code,
once (reference).
Build from source with Go and the Xcode Command Line Tools (on Linux a C compiler and the GL,
X11 and Wayland headers, on Windows MSYS2's MinGW-w64 gcc and make): make run-demo starts a development build on its synthetic demo cluster
(run from source).
| Document | For |
|---|---|
| Usage and reference | Running Cogpit, its flags and files, setting up an agent, cogpit-cli |
| Commands and MCP tools | Every command an agent can call, with its danger level |
| Developer documentation | Architecture, conventions, testing, releasing, decisions |
| Security architecture | Trust boundaries, credentials, privilege footprint, agents and MCP |
| Progress | What exists and what is missing, view by view |
| ROADMAP.md | Where Cogpit stands and what comes next |
| SECURITY.md | Reporting a vulnerability |
| AGENTS.md | Contributing, for people and coding agents |
Apache-2.0. Cogpit is an independent project: it contains no code, text or assets from other Kubernetes IDEs and is not affiliated with their makers. Third-party material, including the documentation method adapted from PhotoCraft, is listed in ATTRIBUTION.md.