SkyPilot: one job interface over Kubernetes, Slurm and 20+ clouds, launched inside your own accounts
The AI Compute Platform for frontier teams. SkyPilot turns fragmented AI compute into one AI supercomputer, so frontier AI teams build custom intelligence faster.
At a glance
- What is it?
- SkyPilot is an Apache-2.0 Python system that runs, queues and recovers AI jobs across Kubernetes clusters, Slurm, reserved GPUs and a long list of cloud providers. This piece walks through how its install extras, bring-your-own-cloud model, Kubernetes layer, utilization features and agent skill actually work, and where each one leaves work for you.
- Who is it for?
- SkyPilot fits teams that already hold cloud accounts, reserved GPUs or Kubernetes and Slurm clusters and want one interface to queue, launch and recover AI jobs across them, plus autostop and binpacking to cut idle capacity. It does not fit anyone looking for a hosted GPU service, because everything runs inside your own accounts and clusters with your own credentials.
- Can I use it commercially?
- Yes. Apache-2.0 is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 2 days ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 13, 2026, and from our analysis. They are not legal advice.
Star history
- skypilot-org/skypilot10,830 ★ · +229
| Repository | 2026-09-13 | 2026-11-13 |
|---|---|---|
| skypilot-org/skypilot | 10.6k | 10.8k |
229 Stars in the last 61 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
The install extras decide which clouds SkyPilot can talk to
SkyPilot ships as a Python package, and the providers it can reach are chosen at install time through extras. The recommended path uses uv:
# Choose your clouds:
uv pip install "skypilot[kubernetes,aws,gcp,azure,oci,nebius,lambda,runpod,fluidstack,paperspace,cudo,ibm,scp,seeweb,shadeform,verda]"The bracket list is the part to edit: the comment above it says to choose your clouds, so the extras should match the providers and clusters you actually hold accounts on, from kubernetes and aws through to seeweb, shadeform and verda. pip, nightly builds and installs from source are also supported, with steps on the installation page at docs.skypilot.co.
The package is built with setuptools; setup.py loads its dependency table from sky/setup_files/dependencies.py through runpy rather than importing it, and it writes the commit hash into sky/__init__.py. If that hash step fails, setup prints a warning that SkyPilot can still be used normally, so a failed hash write is not a broken install.
Bring your own cloud: nothing runs outside your accounts, VPCs and clusters
SkyPilot is BYOC. Every cluster and job it launches lives inside your own cloud accounts, VPCs and clusters, which is the core of how it works and also its main constraint. It does not sell or lend compute. You supply the credentials and quotas, and SkyPilot chooses where to place the workload among them.
The current list of supported infrastructure is long: Kubernetes, Slurm, AWS, GCP, Azure, OCI, CoreWeave, Nebius, Lambda Cloud, RunPod, Fluidstack, Cudo, Digital Ocean, Paperspace, Cloudflare, Samsung, IBM, Vast.ai, VMware vSphere, Seeweb, Prime Intellect, Shadeform, Verda Cloud, VastData and Crusoe. Reserved GPUs sit alongside these as another target.
One inconsistency is worth knowing. The docstring at the top of setup.py still describes SkyPilot as running workloads on AWS, Google Cloud Platform or Microsoft Azure. That text predates the current provider list, so anyone reading package metadata alone will get an outdated picture of what the tool covers.
On Kubernetes it adds scheduling and dev access, but you still bring the clusters
A large share of SkyPilot's pitch targets teams already running Kubernetes. It adds an AI-oriented interface for launching jobs and deployments, and it shortens interactive development on Kubernetes: you can SSH into pods, sync code to them, or connect an IDE, which plain Kubernetes leaves to you.
On the scheduling side it offers gang scheduling, multi-node jobs and queueing, and it can bring several clusters under one control plane, alongside clouds that are not Kubernetes at all. The project describes the result as Slurm-like ease of use with cloud-native behavior, and Slurm clusters are themselves a supported target.
The limit is in the word existing. Every Kubernetes benefit is framed as an upgrade to clusters you already operate, so a team without a cluster gets none of this section until someone stands one up, and SkyPilot's BYOC model means that someone is you. For a feature-by-feature comparison, the project maintains a page called SkyPilot vs Vanilla Kubernetes on docs.skypilot.co, which is the right place to check before assuming a Kubernetes feature is replaced rather than wrapped.
Autostop, binpacking and failover target idle GPUs, not faster kernels
SkyPilot's utilization features are about where and how long hardware runs, not about making a model train faster on a given GPU. Autostop cleans up idle resources automatically, so a forgotten development cluster stops instead of running indefinitely. Binpacking places workloads tightly on shared clusters, and the scheduler sends work to whichever infrastructure is most available.
Provisioning is flexible across GPUs, TPUs and CPUs, with failover: if a requested resource is not available in one place, SkyPilot can try another. Jobs can be queued, run and auto-recovered in bulk, and the project states that existing GPU, TPU and CPU workloads run with no code changes.
The consequence for planning is that these features work only across capacity you have connected. Failover across clouds needs accounts and quotas on more than one cloud, and binpacking needs a shared cluster to pack into. A team with a single account in a single region gets autostop and queueing, but little of the cross-infrastructure scheduling the rest of the design is built around.
Team deployment runs through an API server image with a pinned gcloud
For shared use, SkyPilot offers team deployment and resource sharing through an API server, documented under reference/api-server. The repository carries the pieces for that: a Dockerfile, separate Dockerfile_k8s and Dockerfile_k8s_gpu images, and a charts directory.
The main Dockerfile is multi-stage on python:3.13.15-slim. Its first stage installs the Google Cloud CLI from Google's APT repository at a pinned version, GCLOUD_VERSION=567.0.0-0, and a comment explains why: without a version specifier, registry caching would keep reusing whatever version apt resolved at first build. The pin must stay in sync with _GCLOUD_VERSION in sky/clouds/gcp.py, and 567.0.0 was chosen because it ships gsutil 5.37, needed after pyopenssl 24.3 dropped an API.
A second stage defaults to INSTALL_FROM_SOURCE=true and sets NEXT_BASE_PATH=/dashboard, installing Node.js and npm to build a web dashboard. If you build a custom server image, those two arguments and the gcloud pin are the settings to keep consistent with the source tree you package.
The agent skill installs from HEAD, not from a pinned release
SkyPilot can also be driven by a coding agent such as Claude Code or Codex through the SkyPilot Skill. Installation is a sentence you hand to the agent:
Fetch and follow https://github.com/skypilot-org/skypilot/blob/HEAD/agent/INSTALL.md to install the skypilot skillThe repository has a matching agent directory and a .claude-plugin directory. Note what that URL points at: blob/HEAD, the tip of the default branch, which is master. Whatever INSTALL.md says at the moment the agent fetches it is what the agent will follow, independent of which SkyPilot release you installed with uv. Two teammates who run the same sentence a week apart can end up with different skill setups.
For reproducible setups, read agent/INSTALL.md yourself before handing it over, and compare it with the tag of the release you are running. Letting an agent follow live instructions from a repository also means trusting every future change to that file, on a machine where the agent can reach your cloud credentials.
Three releases in October and features at different stages
Development is fast. The last push landed on 2 November 2026, and releases v0.14.0, v0.14.1 and v0.14.2 shipped on 5, 19 and 28 October 2026. A minor version line that collects three releases in a month is a reason to pin the exact version in requirements and read release notes before moving.
The surface around the core scheduler is also growing in stages. Recent announcements include SkyPilot Endpoints for inference on clusters you own, Agent Sessions for running coding agents on your infrastructure, the GPU Compass dashboard at gpus.skypilot.co for comparing GPU pricing, and SkyPilot Sandboxes for running untrusted, LLM-generated code on Kubernetes. Sandboxes was announced with an early-access sign-up, so it should not be treated as a finished part of the open-source package.
Testing runs in parallel with pytest-xdist using 16 workers and sets SKYPILOT_DISABLE_USAGE_COLLECTION=1 in the test environment, so the test suite runs with usage collection turned off through that environment variable.
Editorial conclusion
SkyPilot fits teams that already hold cloud accounts, reserved GPUs or Kubernetes and Slurm clusters and want one interface to queue, launch and recover AI jobs across them, plus autostop and binpacking to cut idle capacity. It does not fit anyone looking for a hosted GPU service, because everything runs inside your own accounts and clusters with your own credentials. Before adopting it, confirm that the providers you use appear in the install extras, read the SkyPilot vs Vanilla Kubernetes page if you already run Kubernetes, and pin a release rather than tracking master, since three releases shipped in October 2026 alone.
Frequently asked questions
How to install SkyPilot?
Install it with uv and list the clouds you need as extras, for example uv pip install "skypilot[kubernetes,aws,gcp,azure,oci,nebius,lambda,runpod,fluidstack,paperspace,cudo,ibm,scp,seeweb,shadeform,verda]". pip, nightly builds and source installs are also supported.
Is SkyPilot free?
SkyPilot itself is open source under the Apache-2.0 license. It launches everything inside your own cloud accounts, VPCs and clusters, so the compute runs on infrastructure you already pay for.
How does SkyPilot work?
You describe the environment and job as code, and SkyPilot provisions GPUs, TPUs or CPUs on Kubernetes, Slurm, reserved GPUs or one of 20+ clouds, with failover, queueing and auto-recovery. Autostop and binpacking reduce idle resources.
Is SkyPilot open source?
Yes. The skypilot-org/skypilot repository is published under the Apache-2.0 license and written primarily in Python.
SkyPilot vs Kubernetes: does SkyPilot replace it?
No. SkyPilot runs on top of your existing Kubernetes clusters, adding gang scheduling, queueing, multi-cluster control and SSH, code sync and IDE access to pods. The project keeps a SkyPilot vs Vanilla Kubernetes comparison on docs.skypilot.co.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/skypilot-org-skypilot)Star history chart for your README
[](https://hysenlabs.com/projects/skypilot-org-skypilot)Was this useful?