AI automation platform for orchestrating agentic sessions on Kubernetes
The Agent Control Plane (ACP) lets teams create and manage AI agentic sessions - automated tasks that clone repos, run AI agents, and push results. Sessions are stored in PostgreSQL and reconciled into Kubernetes pods via gRPC.
- Agentic Sessions: AI-powered automation for code review, bug fixes, research, and development tasks
- Multi-Agent Workflows: Specialized AI agents with configurable prompts, models, and repos
- Git Provider Support: GitHub and GitLab (SaaS and self-hosted) via credential sidecars
- Kubernetes Execution: Sessions run as pods with RBAC, resource limits, and namespace isolation
- CLI and SDK:
acpctlCLI and generated SDKs (Go, Python, TypeScript) for automation
make kind-up
make kind-port-forward # ports shown in outputSee CONTRIBUTING.md for full local development setup.
When running with OPENSHELL_USE_GATEWAY=true, the control plane delegates sandbox creation to an OpenShell gateway instead of creating pods directly.
make kind-up OPENSHELL_USE_GATEWAY=trueThis automatically installs all prerequisites: the tenant namespace, the agent-sandbox CRD (v0.4.6), and the OpenShell gateway Helm chart. It also patches the control plane deployment with OPENSHELL_USE_GATEWAY=true.
Override defaults with:
| Variable | Default | Description |
|---|---|---|
OPENSHELL_TENANT_NAMESPACE |
tenant |
Namespace for the gateway and sandboxes |
AGENT_SANDBOX_VERSION |
v0.4.6 |
Agent Sandbox CRD release (must match gateway API version) |
After the sandbox reaches Ready, the control plane executes commands inside it via the ExecSandbox gRPC RPC — the runner starts through exec, not the container entrypoint.
See OpenShell Sandbox Provisioning Spec for full details on gateway mode, CRD version compatibility, and configuration.
| Component | Technology | Description |
|---|---|---|
API Server (ambient-api-server) |
Go + rh-trex-ai | REST + gRPC API, PostgreSQL-backed. Source of truth. |
Control Plane (ambient-control-plane) |
Go | Watches API server via gRPC streams, creates K8s pods |
UI (ambient-ui) |
NextJS + Shadcn | Web interface for managing sessions and agents |
Runner (ambient-runner) |
Python | Executes AI agents inside pods (Claude, Gemini, LangGraph bridges) |
MCP Server (ambient-mcp) |
Go | MCP tool definitions, deployed as credential sidecars |
User Creates Session → API Server Persists to DB → Control Plane Creates Pod →
Runner Executes AI Agent → Results Stream to API Server → UI Displays Progress
- Documentation site - user-facing docs (Astro Starlight)
- docs/internal/ - developer and architecture docs
- CLAUDE.md - development standards and conventions
- BOOKMARKS.md - developer reference index
- API Server - REST + gRPC API (rh-trex-ai, PostgreSQL)
- Control Plane - gRPC-driven session reconciler
- UI - NextJS web application
- Runner - AI agent execution
- MCP Server - MCP tool integration
- CLI -
acpctlcommand-line tool - SDK - generated from the OpenAPI spec (Go, Python, TypeScript)
- Manifests - Kustomize deployment resources
See CONTRIBUTING.md for guidelines, code standards, and local development setup.
This project is licensed under the MIT License - see the LICENSE file for details.