Prospero: A new coding harness for a more civilized age
- ID
- 41800
- Status
- summarized
- Published
- 31 Oct 2026, 7:54 PM
- Fetched
- 02 Nov 2026, 4:44 AM
- Provider
- Hacker News
- Category
- dev-community
- Original URL
- https://nullprogram.com/blog/2026/10/31/
- Source URL
- https://hnrss.org/best
Summary
- Score
- 7.0
- Created
- 02 Nov 2026, 4:44 AM
- Tags
- Audience
- developersvibe_codersai_agent_users
What happened
A developer's blog post introduces Prospero, a self-built coding harness written in C++ with a Dear ImGui docking interface, claiming it uses roughly two orders of magnitude fewer resources than other harnesses while keeping agent features: tool loops, subagents that message each other, and headless deployment of itself to other hosts over SSH or WSL. It ships with no model providers built in, so you supply llama.cpp, an on-prem workplace model, OpenRouter, or the Anthropic/OpenAI APIs (the post notes a Claude subscription cannot be used). A paid-tier "Professional Edition" bundles engines for x86/x86-64 and ARM64 across macOS, Windows and Linux, including a Windows x86 engine that runs on Windows XP, and it is one of few harnesses with desktop control: a vision-capable model can click, type and drag to test GUI apps without per-app MCP servers.
Why it matters
If you run coding agents against a local open-weight model on a consumer laptop, the post's core claim is that harness RAM overhead — not the model — is what makes the setup painful, and it offers a C++/ImGui alternative to web-stack harnesses for that case. The concrete constraint to plan around is bring-your-own-provider: no built-in providers and no Claude subscription support means you budget API keys or a local llama.cpp endpoint yourself. Desktop control via vision is the feature worth evaluating against your existing per-app MCP setups, but the post is a self-announcement with no benchmarks, so treat the resource numbers as unverified until you measure.
Discussion angle
The post argues a coding harness is 'essentially curl glued to popen()' and shouldn't need gigabytes of RAM — is web-stack harness overhead a real cost you've felt with local models, and does vision-based desktop control actually replace the per-app MCP servers you maintain?