Prospero is a C++ coding harness for local LLMs whose author claims around two orders of magnitude fewer resources

#ai #coding #cpp #llm

📅 October 31, 2026 • nullprogram.com

A new option for running coding agents on local hardware appeared on October 31, 2026. Prospero is a coding-agent harness written in C++ with a native Dear ImGui interface, designed for local and open-weight models first, while still working with hosted APIs. Its author announced it on nullprogram.com with the claim that it uses around two orders of magnitude fewer resources than other harnesses, without publishing benchmarks to back that up. A coding-agent harness is the layer around a language model that manages tools, files and the conversation while the model writes and edits code.

How it is built

Prospero ships with no built-in providers, which makes bring-your-own-backend the core design principle. The code is published under the Unlicense, which places it in the public domain. Users connect their own llama.cpp setup for fully local work, or point the tool at an API from OpenRouter, Anthropic or OpenAI, and a Claude subscription cannot be used. The same executable can run headless, so the interface is a separate layer rather than part of the core, and the UI can detach and re-attach without killing the running session. The cross-platform headless engines for remote deployment ship with the paid Professional edition starting at 25 dollars, while the free Community build on GitHub has no bundled headless engines, which limits remote deployments there. In practice that means a long coding task can keep running in the background while the window is closed, and the user can return to it later without losing state. This detach-and-reattach behavior is uncommon among agent tools, which usually tie the session to the interface process.

Why resource use matters

The author points out that most harnesses are built with web stacks, with a browser-based interface on top of a runtime stack that can consume a lot of memory before any model work begins. On a machine that also runs a local model, that overhead competes directly with the model weights and the context window, both of which are memory-hungry. A native C++ approach aims to push most of the available memory toward the model itself, which matters most for people running smaller open-weight models on consumer laptops and desktops.

Impact

If the claimed two-orders-of-magnitude reduction holds up in everyday use, it lowers the entry bar for running a coding agent alongside a local model on ordinary hardware. The bring-your-own-provider design also keeps the tool independent of any single vendor, a fit with the open-weight ecosystem it targets. For users of larger models through APIs, the benefit is smaller, since the harness overhead is a fixed cost on top of network latency and provider pricing. Whether the resource claim survives contact with real workloads will be the first thing the community tests.

Sources: nullprogram.com and github.com