Chris Wellons has been building tools his way for years — tight C/C++, no unnecessary dependencies, respect for the machine. Prospero is his answer to the current generation of AI coding harnesses (Cursor, Aider, Continue, and their kin), which he finds inexplicably bloated for what amounts to an HTTP client glued to a process spawner. His harness uses Dear ImGui for its interface, runs in roughly two orders of magnitude less memory than competitors, and ships with no model provider baked in — you bring your own llama.cpp instance, API key, or on-prem endpoint. The feature set is not stripped down. Prospero supports subagent spawning, inter-agent messaging, headless deployment across hosts (with engines for x86, x86-64, and ARM64 on macOS, Windows, and Linux), desktop vision and control, Git worktrees, MCP protocol, web search, and session forking. The Windows x86 engine runs on XP. The Professional Edition embeds cross-platform engines so agents can operate on remote systems via SSH or WSL. Two design principles define Prospero's character. First: never wait forever. Every operation has a deadline. Shell commands that exceed their timeout convert to asynchronous background tasks rather than hanging the session. If you're away and a permission request goes unanswered, the agent gets a denial and adapts. System toasts alert you when attention is needed. This is a direct response to a real workflow problem — setting a long task overnight and finding the harness stuck on a trivial prompt two hours in. Second: no stray processes. The entire process tree lives and dies together. When a command is stopped, it stops — its children don't linger consuming resources and interfering with subsequent runs. Wellons notes this is trivial on Windows but harder on macOS and Linux; Prospero follows through on all three. Combined with the lean footprint, this makes Prospero possibly the only harness that can reliably drive a fuzz campaign without the Linux OOM killer targeting the harness instead of the fuzz target. Configuration is INI-based (not JSON, not YAML), with URL-based includes for shared or managed configs. The harness fetches model lists and context sizes from endpoints automatically. You can switch shells mid-session — PowerShell to bash to zsh — and the agent is informed of the change and its quirks. A built-in skill lets agents read Prospero's own documentation, answer configuration questions with hyperlinks to the relevant UI, and edit the INI on your behalf. The permissions model is refreshingly honest: it will not stop a malicious agent. Agents write and run programs on your computer. Permissions are gentle reminders and mistake-catchers, not security boundaries. If you want real isolation, you containerize the harness or put it on a Raspberry Pi. No security theater. Prospero is a statement piece as much as a tool. It argues that the coding harness problem is fundamentally simple — HTTP calls and process management — and that the industry's default stack choices (Electron, Node, web frameworks) impose costs wildly disproportionate to the task. Whether that argument scales beyond a single deeply skilled developer's workflow is the open question.
