GitHub Copilot SDK

The GitHub Copilot SDK takes the agentic execution loop that powers Copilot CLI and packages it as a multi-platform library you can call from your own code. This is not a chat widget or a thin LLM wrapper — you get the full planning, tool-invocation, file-editing, multi-turn-session, and streaming runtime, with MCP server support baked in.

It went generally available on June 2, 2026 with six language bindings at launch: Node.js/TypeScript, Python, Go, .NET, Rust, and Java. You do not need a Copilot subscription to use it — BYOK support lets you bring your own API key from OpenAI, Anthropic, Azure, AWS Bedrock, Google AI Studio, or any OpenAI-compatible provider, and usage bills to that provider. The SDK itself is MIT-licensed.

The honest caveat: embedding an agent engine means you inherit the governance problem — tool scoping, approval hooks, sandbox policy, and audit logging are your job, and the official sandboxes are still in preview. Reach for this when you want a production-tested agent runtime without building your own planner and tool loop; skip it if all you need is a completion API.

1. Install the SDK for your language (e.g. npm install @github/copilot-sdk or the equivalent package).
2. Authenticate: use a Copilot subscription token, or configure BYOK with your own provider API key.
3. Define your domain tools and MCP server configuration, including tool filtering and approval policies.
4. Start a session, stream events, and wire the agent loop into your app's workflow.