# GitHub opens Copilot SDK public preview in five languages > The preview exposes Copilot’s agent runtime for application builders, including tools, streaming, approvals, tracing, attachments, and BYOK. Canonical URL: https://www.devobs.io/news/news-github-copilot-sdk-public-preview/ By: Arjun Shah Published: 2026-09-06T11:58:54.641Z Updated: 2026-09-06T11:58:54.641Z Event date: 2026-04-02 Section: Developer tools GitHub put the Copilot SDK into public preview on April 2, offering application developers the agent runtime used by Copilot cloud agent and Copilot CLI. The [announcement](https://github.blog/changelog/2026-04-02-copilot-sdk-in-public-preview/) lists packages for Node.js and TypeScript, Python, Go, .NET, and Java, with tool invocation, streaming, file operations, and multi-turn sessions. ## An agent runtime becomes an SDK The release is broader than a text-generation wrapper. Developers can define custom tools and handlers, customize portions of a system prompt, attach binary data, and gate sensitive operations through approval handlers. Built-in OpenTelemetry support uses W3C trace context, which gives teams a starting point for connecting agent activity to existing observability systems. Bring-your-own-key options are also described for several model providers. Public preview is the key qualifier. Interfaces, quotas, and provider support can change, so a production design should isolate the SDK behind an application-owned boundary. It should also preserve a deterministic path for operations that cannot tolerate model variability. ## Review permissions before capability The first design exercise should be a tool matrix: what each tool can read, what it can change, which identity it uses, and when a person must approve it. Marking a tool read-only is meaningful only if the handler truly cannot mutate external state. Validate arguments again inside the handler rather than trusting model-generated input. Instrument session IDs, tool decisions, latency, and failure categories while excluding prompts or attachments that contain secrets. Exercise cancellation and partial-failure behavior, because a streamed answer may stop after an external action has already completed. For an initial trial, choose a bounded workflow with reversible effects and a clear success check. That will reveal integration quality more reliably than a polished chat demonstration. ## Source references -