What is Kyber?
Kyber is a self-hosted platform for running and managing a fleet of long-lived AI agents on a Kubernetes cluster. Each agent gets full autonomous access to its own sandboxed pod, including disk, tools, repos, and memory, and keeps all of it across restarts. You manage the fleet from a web console, the API, or from Telegram, Discord, and Slack on your phone. It is made to run with the Claude or ChatGPT subscription you already pay for.
The mental model
Section titled “The mental model”An agent in Kyber is not a chat session. It is long-lived infrastructure: a worker with its own sandboxed pod, its own persistent filesystem, its own identity, and its own model. Installed packages, cloned repos, credentials, and memory survive pod restarts and upgrades. Recovery after machine loss depends on the storage backend: portable volumes can reattach, while node-local volumes are tied to their node.
You declare what you want, an agent of a certain shape on a certain kind of machine, and Kyber works to make reality match that intent and keep it there. Cron inside each agent survives restarts, and agents can send each other signed messages, so one agent’s output becomes another’s next prompt.
Agents and machines
Section titled “Agents and machines”Two concepts an operator creates and manages:
- Agent: a single AI agent instance. You choose its model, its compute size, its identity, and which machine it targets. Its filesystem is durable, so its work and memory persist across pod recreation, restarts, and machine interruption. An agent can also be given a private identity repo, its own GitHub repository of persona, long-term memory, and session state, so its identity is versioned and survives even a full teardown. See Memory and identity.
- Machine: where agents run. A Machine can be a cloud VM that Kyber provisions and manages for you (including direct GCE, managed GKE pools, and managed EKS node groups with installer-curated profiles), or it can stand for a node that already exists in your cluster, the shape used by single-box installs and by managed Kubernetes services where your own node pool provides the capacity. Kyber handles bringing machines up and recovering agents when interruptible capacity is reclaimed.
What ships in the box
Section titled “What ships in the box”- A control plane that manages agents and machines, plus in-cluster Postgres and Redis. Everything installs from a published Helm chart.
- The fleet console, a browser PWA served by the control plane: agent status, per-agent token usage and context budget, a live terminal into any running agent.
- Two agent runtimes: Claude Code and Codex. Claude Code agents sign in with your Claude subscription, Codex agents with your ChatGPT subscription. API keys work too. See Runtimes.
- Telegram, Discord, and Slack channels for two-way chat with any agent, no terminal needed. See Chat channels.
Who it is for
Section titled “Who it is for”Kyber is for operators who want to own the whole environment, not just the agent: pick each agent’s machine, disk, CPU, and memory, reboot it, stop it, or open a shell into it. It is self-hosted, licensed Apache 2.0, at v1.x, solo-maintained, and runs the maintainer’s own fleets every day.
Ready to try it? Start with the quickstart or browse the install options.