github/copilot-sdk
AI EngineeringEmbed GitHub Copilot's agent loop in your own app without building orchestration, for teams with Copilot access or their own model keys.
This is worth a look if you already want Copilot's agent behaviour inside a product, but it is not a drop-in component: it wraps the agent engine behind Copilot CLI, so its practical weight is an external runtime plus your choice of auth and billing model.
The problem
Suppose you want an app that can plan a task, call tools, and edit files, but you do not want to write the loop that sequences model turns, tool calls and results. Before this SDK, you would either build that orchestration yourself or drive Copilot CLI in an ad-hoc way. With the SDK, you define the agent's behaviour and tool surface in your language of choice, and the SDK invokes the same engine behind Copilot CLI to handle planning and tool invocation. The result is a programmatic agent in your app rather than a separate CLI session you have to script around.
What's different
The distinctive part is that the SDK exposes an existing agent runtime rather than a new orchestration framework. It ships first-party packages for Python, TypeScript, Go, .NET, Java and Rust, so the same engine can sit behind services written in quite different stacks. It also supports several authentication paths: a signed-in GitHub user, an OAuth GitHub App, environment tokens, or BYOK with provider API keys. That flexibility matters because it separates the runtime from how you pay and authenticate.
The main trade-off is coupling. You get Copilot's behaviour and defaults, not a blank model interface with your own planner. Billing follows the Copilot CLI model, with each prompt counting toward a usage allowance, which is a different cost shape from paying a provider directly. BYOK avoids GitHub auth but is key-based only; Microsoft Entra ID, managed identities and third-party identity providers are not supported. The skill installation APIs also show that the SDK carries a wider surface than a simple chat call, and that surface needs to be managed rather than ignored.
Who should try it
A good fit is a team building an internal or customer-facing assistant that needs to take actions in files or tools, already has a Copilot plan, and wants to avoid writing the planner-tool loop. A poor fit is a team that needs full control of orchestration, wants an identity provider such as Entra ID for model access, or expects a self-contained library with no external runtime. For Go, Java and Rust, the Copilot CLI is not bundled by default and must be installed manually or present on PATH; Go and Rust do offer application-level bundling, while Node.js, Python and .NET can rely on the bundled CLI. That difference changes both deployment and who can use the SDK comfortably.
Requirements and limitations
Plan for the CLI runtime as an ongoing dependency. JavaScript, Python and .NET users get it bundled, but Go, Java and Rust deployments need the CLI installed or bundled explicitly. Authentication is either GitHub-based or BYOK; check the auth documentation and BYOK setup before committing, because BYOK excludes Entra ID and managed identities. Billing is per-prompt against a Copilot allowance, so usage-based cost planning is required rather than simple model-token budgeting. If you plan to expose skill installation, the installation APIs involve planning, applying, recovery, status and enable/disable states, meaning your UI needs a review and recovery flow, not a one-shot install button.
Start from the getting started guide and the per-language package pages for Node.js, Python, .NET, Go, Rust and Java. The README is selected excerpts, so verify language-specific deployment details there rather than assuming uniform bundling.
HotRepo's take
The repository was created on 2026-01-14 and recorded 10,586 stars on 2026-10-28. The supplied board sample starts on 2026-07-16, and the final recorded day was only +5 stars, so interest exists but persistence cannot be judged from these nine observed dates. Adoption is reasonable for teams already inside the Copilot ecosystem; confirm the CLI bundling and auth model for your chosen language before building around it.
Written by AI from the project's README, updated October 28, 2026. The project's own docs are the source of truth.