Kelos

Kelos is an open-source Kubernetes-native framework for orchestrating autonomous AI coding agents. It lets you point an agent at a GitHub issue and get a pull request back — fully autonomous, running in isolated ephemeral Pods on your Kubernetes cluster. Kelos supports Claude Code, OpenAI Codex, Google Gemini, OpenCode, and custom agent images through a standardized container interface. Written in Go, it installs as a Kubernetes operator with custom resources and ships with a dedicated CLI.

Kelos treats agent runs as first-class Kubernetes resources with four core primitives: Tasks (units of work wrapping an agent run), Workspaces (git repos where agents operate), AgentConfigs (reusable bundles of instructions, plugins, and MCP servers), and TaskSpawners (event-driven engines that react to GitHub issues, PRs, or cron schedules to create tasks automatically).

• Isolated Ephemeral Pods - Each agent task runs in its own Kubernetes Pod with a freshly cloned git workspace and no access to your host machine.
• Task Pipelines - Chain tasks with dependsOn and pass results like branch names, PR URLs, and token usage between pipeline stages.
• TaskSpawner - Automatically create agent tasks in response to GitHub issues, pull requests, or cron schedules with configurable polling and concurrency limits.
• Multi-Agent Support - Run Claude Code, OpenAI Codex, Google Gemini, OpenCode, or any custom agent image using a simple container interface.
• AgentConfig - Bundle project-wide instructions, plugins, skills, and MCP server configurations into reusable configs that get injected into agent Pods.
• CLI and Declarative YAML - Manage tasks via the kelos CLI or apply Kubernetes YAML manifests directly, making it GitOps-ready and compatible with ArgoCD.
• Autonomous Self-Development - Build feedback loops where agents pick up issues, write code, self-review, fix CI flakes, and label issues needing human input.
• Scalable Parallelism - Fan out agents across multiple repositories with Kubernetes handling scheduling, resource management, and queueing.
• Observable Outputs - Every run captures deterministic outputs including branch names, commit SHAs, PR URLs, and token usage into Kubernetes resource status.
• Cost Controls - Set maxConcurrency, maxTotalTasks, and per-task timeouts to prevent runaway agent activity and control API spend.