GitHub released its Copilot SDK to technical preview on January 14, exposing the production agent runtime that powers Copilot CLI. The SDK covers TypeScript, Python, Go, and .NET, letting developers add multi-turn conversations, tool execution, and context management to their apps without building custom orchestration layers.
What's notable: The SDK handles tool calling automatically. A demo app called Flight Diary shows this in practice - users upload boarding pass images, and Copilot extracts flight details (no separate OCR service), then queries Azure Cosmos DB using natural language (no SQL expertise needed). The agent decides when to parse the image and when to save data based on the prompt context.
The architecture uses JSON-RPC to access Copilot CLI's capabilities - Git operations, file edits, web requests, and planning. Tools run with --allow-all permissions by default (customizable, but file system access warrants attention). Requires a GitHub Copilot subscription; Microsoft's Work IQ features need Microsoft 365 Copilot and tenant approval.
This matters for enterprise teams evaluating agent frameworks. GitHub is positioning this as "agents for every app" infrastructure rather than a specialty tool. The SDK supports all CLI models, queryable at runtime, and works with local development environments (Cosmos DB vNext emulator in the demo).
The trade-offs: It's early. Technical preview means CLI dependency and evolving APIs. GitHub wants feedback via their Community forum, which signals they're still shaping capabilities based on real implementations.
Worth watching how this compares to building on foundation models directly. If the abstraction holds up, it could simplify agent deployment for teams already in the GitHub ecosystem. If it doesn't, you've coupled your workflow to GitHub's runtime choices.
Demo code available on GitHub. The Flight Diary example shows state management and context handling patterns that transfer to other use cases.