Built to be trusted with production.
Everything an operations team asks before pointing a tool at a live cluster — answered by how SKYLINE is built, not by a policy page.
Read-only by default
Each environment has an access level: read-only, write or full. The level is a command allow-list enforced by the runner, not a request in a prompt. Writes are explicit grants, per environment.
Every command on record
Each step and its output stays with the investigation; failed and refused commands too. Administrative actions land in an audit log.
Credentials stay on the server
Kubeconfigs and data-source credentials are encrypted at rest; the browser never sees them. The in-cluster agent keeps its own secret in your cluster and signs every request.
Per-customer isolation
Projects, environments, documents, connections and models are scoped by group. Queries to metrics and logs are pinned to the environment they belong to.
Deploy it your way
SaaS, private SaaS or hosted in your own environment. SKYLINE maintains it in every case, and the agent in your network only connects out.
Your models, your keys
Use a model SKYLINE provides, any OpenAI-compatible provider or a model you host yourself. Which model runs an investigation is your choice, per workspace.
Where your data is processed.
It depends on the deployment you choose. If your data has to stay in your environment, choose Private SaaS or Customer-hosted.
How far SKYLINE may act.
Each environment has its own limit, so production can stay read-only while staging allows approved changes.
Investigate and recommend
The read-only default. SKYLINE finds the cause and hands you the fix as commands; you run them through your normal change process.
Change with approval
On a write or full environment, each change waits for a person to approve it. Project rules can allow chosen commands, only while someone is watching.
Change on its own
Not offered. Runs nobody is watching, such as alert investigations, cannot change anything.
The details your reviewers will ask about.
How access levels map to commands, how the agent connects and what it may run: Access levels and Add an environment in the docs.
- Egress allow-listing. SKYLINE connects to your endpoints from fixed addresses, published for your firewall rules.
- Tools off by default. An MCP server's tools start disabled and are enabled one by one; tools that change state can require approval.
- Agent and bridge. Outbound-only connections with HMAC-signed requests; an unsigned or stale agent is refused.
- Signed releases. The agent image, its Helm chart and the bridge binaries are signed with cosign by our release pipeline. Check them with our public key at skylinehq.io/cosign.pub before you install.
- Maintenance and limits. Per-workspace limits and timeouts, a maintenance mode, and usage you can inspect per model.
Questions for your security review?
Request access and mention it; we walk through the deployment with your team.
Request access