ferroxlabs/wayland-core

Wayland Core is an AI agent engine written in Rust that runs from the command line. You give it a task in plain text, and it connects to a language model, uses local tools like file reading and shell commands, and works through the task step by step. It is designed to be a standalone engine, not tied to any single AI provider or any particular editor. The tool supports around 20 AI providers out of the box, including Anthropic, OpenAI, Google Gemini, AWS, Azure, Groq, DeepSeek, Mistral, and others, plus a catalog of over a hundred additional model endpoints. Switching between providers is a configuration change rather than a code change. When a provider call fails, the engine automatically retries, reconnects if the connection drops mid-stream, rotates through backup API keys, and falls over to a different provider if one is set as a fallback. Beyond running a single agent, Wayland Core can fan work out to multiple agents running in parallel. A worktree swarm runs several workers simultaneously, each in its own isolated copy of the repository on its own branch, so a failure in one worker does not affect the others. Declarative workflows let you define a sequence of stages that are validated against a schema, and a stage that produces invalid output is automatically retried with the error included. Tool execution happens inside a sandbox by default. On Linux it uses bubblewrap, on macOS it uses sandbox-exec, and on Windows it uses AppContainer. If no working sandbox is available and the user has not explicitly opted out, the engine refuses to run model-driven shell commands rather than running them without protection. Network egress is controlled through a single point, and a classifier blocks requests that look like they might be sending data to unexpected destinations. The engine runs three ways from one binary: as a one-shot command that answers a question and exits, as a full-screen interactive terminal interface, or as a headless process that another application drives over a JSON-Lines protocol. It is available through npm as a prebuilt binary, from GitHub Releases, or compiled from source. The project is licensed under Apache-2.0.