Instead of a kubeconfig per developer, per cluster
No credentials on laptops
Users sign in to the control plane. Cluster credentials never leave the agent, and access is revoked in one place.
What changes
Giving an engineer access usually means handing out a credential or opening a network path. kbridge puts a control plane between people and clusters, so you do neither.
Instead of a kubeconfig per developer, per cluster
Users sign in to the control plane. Cluster credentials never leave the agent, and access is revoked in one place.
Instead of a VPN, a bastion or a public endpoint
Each cluster's agent dials out to the control plane, so there is nothing inbound to open or maintain.
Instead of a shared credential that hides who did what
The agent impersonates kbridge:<email>, so Kubernetes RBAC and the cluster's own audit log see the user, not the agent.
Instead of stitching together logs from several clusters
Who ran which command where, and whether it ran. Refused commands are recorded too.
Automation and AI agents
An automation or AI agent with its own kbridge account holds no cluster credential. Every command it runs is checked against its role and recorded under its name, so you can see exactly what it did and what it was refused.
How it works
It dials out to the control plane over gRPC. No inbound port is opened and no kubeconfig leaves the cluster.
The CLI sends their login token and the command to the control plane. There is nothing new to learn.
The user's role decides. Nothing reaches a cluster without passing this one check, and a refusal says why and who can help.
It impersonates kbridge:<email>, so the cluster's RBAC and audit log see the person, not the agent.
Whether it ran or was refused: user, cluster, command and outcome, in the kbridge audit log.
Admin console
A read-only web console ships inside the control plane. It makes no third-party requests.
Security in brief
Only the agent inside the cluster does, and it never hands it out.
Through impersonation, so Kubernetes RBAC still applies and the cluster's audit log names the person.
Allowed or refused, in one audit log across every cluster.
Coming next
Not in v0.3.1-alpha.1 yet. Each card says how far along it is.
In development
Rules that evaluate the command itself before it is forwarded, so a destructive operation is refused or needs a reason.
In development
Request a role for a bounded window with a reason. Someone approves, and it expires on its own.
Planned
OIDC login, so people sign in with the identity provider they already have.
FAQ
kbridge is an open-source, self-hosted access gateway for Kubernetes. Users run kubectl commands through the kb CLI without holding cluster credentials. A control plane checks each command against the user's role, an agent inside the cluster runs it as that user, and every attempt is recorded.
Create a kbridge account for them and bind it to a role. They install kb, log in, and run kubectl commands. No kubeconfig is created or distributed: the cluster credentials stay with the agent inside the cluster.
No. The agent in each cluster opens an outbound connection to the control plane, so nothing inbound is opened on the cluster. Users only need to reach the control plane.
Every attempt is recorded with the user, cluster, namespace, command and outcome, whether it ran or was refused. Admins read it with kb admin audit or in the admin console. Commands run as the user, so the cluster's own audit log names the person too.
Yes. The agent runs each command as kbridge:<email> through impersonation, so Kubernetes RBAC applies to every request. The agent also keeps the cluster's RoleBindings in step with the kbridge policy.
Yes. Give each one its own kbridge account. It holds no cluster credential, every command it runs is checked against its role, and the audit log records what it did under its own name.
kb takes kubectl syntax, including streaming commands such as logs -f, exec -it and port-forward. There is nothing new to learn.
Yes. kbridge is open source under the Apache 2.0 license and self-hosted: one control plane and one small agent per cluster, installed with Helm.
Start here
The Helm quick start puts the control plane and an agent on one cluster. It works on a local kind cluster.
Helm quick start →Secrets, TLS and more clusters. Then the security checklist, onboarding users, and operations.
Production install →Your admin gave you a URL, an email and a password. Install kb, log in, and run commands.
CLI reference →