GitHub has announced the GitHub Copilot SDK, a technical preview SDK that allows integrating the same agent engine used by Copilot CLI directly into custom applications. It is not a traditional chat API, but a complete and programmable agent runtime.
The SDK exposes the core execution engine of GitHub Copilot CLI: multi-step planning, context management across turns, tool invocation, file editing, command execution, and real-time streaming. All of this without requiring developers to build their own agent loop, tool orchestrator, or memory system. That complexity is already handled.
The architecture is straightforward: your application (web, desktop, backend, or internal tool) communicates with the SDK, which in turn communicates with Copilot CLI running in server mode via JSON-RPC. The SDK manages the process lifecycle, selection of available Copilot models, GitHub authentication, and integration with tools and MCP servers.
From a practical standpoint, this opens the door to building products that previously required significant infrastructure. For example: internal enterprise tools capable of acting on real systems, developer assistants that modify code or execute complete workflows, backoffices and admin panels operable via natural language, automation of complex processes, custom GUIs for agents (similar to Claude Desktop, but embedded in your product), or specialized MCP orchestrators. It also enables the creation of vertical copilots focused on specific domains such as legal, operations, data, or infrastructure.
The SDK requires a GitHub Copilot subscription, depends on Copilot CLI, and is primarily aimed at internal tools or controlled environments rather than large-scale B2C chatbots. Even so, as an agent execution platform, the approach is solid: GitHub is not just offering models, but an operating system for agents that can be reused in any application.
Here you can find the GitHub Copilot CLI SDK
