Skip to content

Agent Control Plane

Agent Control Plane (ACP) runs AI agents against real developer work: repositories, project context, credentials, messages, and long-running tasks. You can start work from the web UI, acpctl, the REST API, generated SDKs, or the MCP server.

ACP is not a CRD data model. The API server is a Go service backed by PostgreSQL and exposes /api/ambient/v1/... on port 8000. The control plane watches the API server over gRPC and creates Kubernetes runner Pods for sessions. Each runner is a Python AG-UI server with bridges for Claude Agent SDK, Gemini CLI, and LangGraph.

  • Running a code agent on a repository with project instructions and credentials.
  • Creating reusable project agents with standing prompts.
  • Starting sessions from the UI, CLI, CI, scripts, SDKs, or MCP tools.
  • Streaming messages and AG-UI events while the agent works.
  • Inspecting files, Git state, repository status, artifacts, and MCP status for active sessions.
  1. Read What is ACP? for the system model.
  2. Use Quick start to create a project, agent, and session from the UI.
  3. Use CLI Reference when you want repeatable automation from a terminal or CI job.
  • Projects group agents, sessions, credentials, and settings.
  • Sessions are the execution records that the control plane turns into runner Pods.
  • Credentials are credential records and sidecars that let runners reach GitHub, GitLab, Jira, Google, Vertex, and Kubernetes.
  • Context & artifacts explains what the runner sees in /workspace and what you can retrieve afterward.
  • Workflows are optional Git-backed instruction bundles loaded into a session.
  • Scheduled sessions are project records for recurring work; the current API stores and manages them, but automatic firing is not implemented in the API service yet.
  • GitHub Actions can create or start sessions with curl or acpctl.
  • MCP Server exposes ACP project, agent, session, and message tools to MCP clients.
  • Bugfix, triage, and PRD/RFE pages give prompt patterns you can run as ordinary ACP sessions.