Prospero: A new coding harness for a more civilized age

Prospero is a lightweight coding harness for local and hosted language models, built to avoid the memory-heavy web stacks common in competing tools. It presents an ImGui interface while letting AI agents use tools such as file operations and shell commands, delegate work to subagents, and coordinate across machines. Vision-capable models can also control the desktop to test graphical applications. The author argues that a lean harness matters especially for long-running work such as fuzzing, where memory-hungry software may be killed before it can finish. Prospero has no model provider built in; users connect their own service or local inference engine.

Its design centers on bounded operations: network calls, permissions, and agent mailboxes have deadlines, while timed-out shell commands become background tasks that agents can revisit. Processes are managed as a tree so stopping a command also stops its descendants. Other features include passive rendering to avoid needless GPU use, editable INI configuration with controlled remote includes, switchable shells, MCP, web search, Git worktrees, and built-in guidance that can link to settings or change configuration. There is no embedded terminal or editor, and permissions are not a security boundary; isolation requires a container or separate machine. The free Community Edition lacks built-in remote engines, while the paid Professional Edition adds headless engines for remote deployment and sessions.