GitHub exposed the SDK for its internal agent runtime, which powers GitHub Copilot. The GitHub SDK is a programmable toolkit. Copilot-SDKYou can now embed this agentic execution loop in any application. This allows the agent to execute commands, plan workflows, run tools and edit files.
What GitHub Copilot provides
The GitHub SDK-SDK integrates the GitHub agent into services and applications. The GitHub Copilot-SDK gives you programmatic access the execution loop which powers GitHub Copilot’s CLI. Instead of building a planner and tool loop, for each project you can use this runtime as an execution system.
Copilot SDK on GitHub exposes the exact same runtime as Copilot CLI. It includes support for Multi-Model Operation, Multi Step Planning, Tools, Model Context Protocol integration (MCP), authentication and streaming. The same behavior as Copilot’s terminal agent is available in your code.
The agentic loop of execution as a primitive runtime
Agentic execution is at the core of this abstraction. The interactions in Copilot SDK and CLI are not discrete prompts. The agent is able to maintain state throughout turns. It chooses tools, performs commands and reads their results.
The GitHub Team describes how to solve the common problems you encounter when implementing this loop. The GitHub team describes the usual problems when you implement this loop yourself.
Core API and supported languages
This technical preview is available in four languages:
- The package includes Node.js, TypeScript and TypeScript.
@github/copilot-cli-sdk - Python through the package
copilot - The module is a good place to start.
github.com/github/copilot-cli-sdk-go - The package.NET
GitHub.Copilot.SDK
The API surfaces of all SDKs are consistent. According to the Changelog, each language binding provides multi-turn sessions with history and custom tool execution. It also allows programmatic control of client and session lifetime cycles.
Toolkits, MCP servers and integration of existing systems
The Copilot agent’s main feature is the execution of tools. You can create custom tools through the SDK that can be called by the model during a chat. Copilot-CLI exposes already custom tool definitions, as well as full MCP integration. The SDK reuses this capability.
MCP provides a standard protocol that allows agents to communicate with external systems, such as APIs or document storage, and operations tools. The Copilot Agent can call and discover its MCP operations using consistent metadata when you integrate the MCP server.
Pattern is easy to understand. It is straightforward.
Subscribe to a streaming service, authenticate, and get a subscription
SDK supports GitHub subscriptions and authentication. When configuring the SDK, you can use either an existing GitHub Copilot membership or your own key. It is crucial to embed the agent into enterprise environments, where identity management and access controls are already standardised around GitHub.
Contracts include streaming. The SDK exposes the streaming functionality so applications can get incremental updates. You can build user interfaces which update incrementally as the agent executes and reasons, rather than waiting until it is complete.
Relationship with the GitHub Copilot CLI
It is not an agent. The SDK is an overlay on the Copilot CLI. It is a method to reuse CLI behavior such as planning, using tools, and multiturn execution in different environments.
Copilot’s CLI continues to grow. Recently, updates have added persistent memory and infinite sessions. Support for plan and explore workflows is now available with model selection per stage. Custom agents and skills are also supported, as well as full support of MCP and asynchronous task delegating. SDK also benefits because the same behaviour is exposed through specific language libraries.
The Key Takeaways
- GitHub Copilot SDK provides the same agentic loop as GitHub Copilot CLI. Applications can then call a tested production planner to run multi-step workflows using tools and commands.
- SDKs for Node.js and Python are also available.
- The SDK allows developers to define Model Context Protocol and domain-specific tools. They then register these servers and Copilot decides how and when they should be called as part of a plan.
- The runtime is integrated with GitHub and Copilot subscriptions. Supports multiple AI models including GPT based backends. Exposes real-time stream so applications can render incremental responses.
Take a look at the GitHub Page. Also, feel free to follow us on Twitter Join our Facebook group! 100k+ ML SubReddit Subscribe now our Newsletter. Wait! Are you using Telegram? now you can join us on telegram as well.

