---
type: "post"
id: "chris-wellons-releases-prospero-a-native-c-dear-09da4f"
slug: "chris-wellons-releases-prospero-a-native-c-dear-09da4f"
title: "Chris Wellons releases Prospero, a native C++ Dear ImGui coding-agent harness using about 100x fewer resources"
description: "Chris Wellons has released Prospero, a native C++ coding-agent harness with a Dear ImGui interface and a design focused on low resource use and bounded operations."
retrieval_nugget: "Chris Wellons has released Prospero, a native C++ coding-agent harness with a Dear ImGui interface and a design focused on low resource use and bounded operations. The source claims roughly 100-fold lower resource consumption than web-stack harnesses; its features make it a practical alternative to evaluate for long-running or locally hosted agent work."
published_at: "2026-11-01"
updated_at: "2026-11-01"
record_date: "2026-11-01"
date_kind: "published_at"
topics: ["anthropic"]
entities: ["Prospero","Chris Wellons","Dear ImGui","OpenRouter","Anthropic","OpenAI","llama.cpp"]
related_patterns: ["harness-architecture-outlives-model-choice"]
novelty_label: "new_fact"
what_changed: "Chris Wellons releases Prospero, a distinct native C++ Dear ImGui coding-agent harness claiming about 100x lower resource use; no prior document covers this release."
prior_refs: []
source_url: "https://nullprogram.com/blog/2026/10/31"
source_urls: ["https://nullprogram.com/blog/2026/10/31"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:chris-wellons-releases-prospero-a-native-c-dear-09da4f"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/chris-wellons-releases-prospero-a-native-c-dear-09da4f/","markdown":"https://newruntime.com/posts/chris-wellons-releases-prospero-a-native-c-dear-09da4f.md","json":"https://newruntime.com/posts/chris-wellons-releases-prospero-a-native-c-dear-09da4f.json"}
---

# Chris Wellons releases Prospero, a native C++ Dear ImGui coding-agent harness using about 100x fewer resources

## Retrieval answer

Chris Wellons has released Prospero, a native C++ coding-agent harness with a Dear ImGui interface and a design focused on low resource use and bounded operations. The source claims roughly 100-fold lower resource consumption than web-stack harnesses; its features make it a practical alternative to evaluate for long-running or locally hosted agent work.

## Summary
Chris Wellons has released Prospero, a native C++ coding-agent harness with a Dear ImGui interface and a design focused on low resource use and bounded operations. The source claims roughly 100-fold lower resource consumption than web-stack harnesses; its features make it a practical alternative to evaluate for long-running or locally hosted agent work.

## What happened
In an October 31, 2026 post, Wellons introduces Prospero as his answer to coding harnesses that, in his view, consume excessive memory despite doing work that can be understood as connecting a language model to tools. Prospero is written in C++ and uses a Dear ImGui docking interface. He says it uses about two orders of magnitude fewer resources than other harnesses without sacrificing features. The post does not define the resource metric or provide a comparative test, so the figure should be read as the author's claim rather than an independently established result.

Prospero is a harness, not a model provider. It has no model providers built in; users configure an endpoint themselves. The post names llama.cpp, an on-premises model, OpenRouter, Anthropic API and OpenAI API as possible sources. The author explicitly says that a Claude subscription cannot be used. The source describes a coding harness as an agent loop in which the model calls tools such as file operations or shell commands and receives their outputs. Prospero also supports subagents that can work in parallel or keep separate context.

The post describes both desktop and remote operation. Its Professional Edition embeds engines for x86/x86-64 and ARM64 on macOS, Windows and Linux. A headless copy can be deployed to another host over SSH or plain WSL; the interface can later reattach to a harness running independently of the GUI machine. The author also describes desktop-control capabilities: a vision-capable model can see and interact with a desktop to test or debug GUI applications. The source says this avoids the need for specialized per-app MCP assistance. The Community Edition is available on GitHub under the Unlicense, and the Professional Edition is sold on itch.io; the post does not give a price.

Several operational choices target unattended work. Every operation has a deadline, network requests have timeouts, watchdogs prevent indefinite waits, and agent mailboxes are bounded. Permission requests time out as denials if the user does not respond. Commands that exceed their timeout become asynchronous background commands, and Prospero aims to stop the entire process tree when a command is stopped. The author argues this matters for builds and fuzzing, where orphaned processes or a harness killed under memory pressure can derail work. Configuration is stored in an INI file, with URL includes restricted from setting values that could appear in shell commands; the author notes that configuration can still expose environment variables as API keys to a configured endpoint.

## What the archive adds
The archive has covered [DeepSeek's research-prototype coding agent for DeepSeek Harness](https://newruntime.com/posts/deepseek-open-sources-dsh-superagent-a-research-prototype-524cb5/) and [RL-trained adapters used with the pi coding agent](https://newruntime.com/posts/harness-hub-releases-rl-trained-lora-adapters-that-2c32d2/). Those are adjacent harness and coding-agent developments, not Prospero coverage. This is a new event in the broader thread of alternative coding harnesses.

## Trend signal
Prospero supports [harness architecture outlives model choice](https://newruntime.com/patterns/harness-architecture-outlives-model-choice/): it is built to connect to multiple external model endpoints, while its distinguishing claims concern the runtime, interface, tool loop and process behavior. The release is a concrete example, not evidence that its resource claim or design will generalize across harnesses.

## Why now
Wellons frames the release around the increasing usefulness of open-weight models on higher-end consumer hardware for simpler programming tasks. His stated motivation is that such models need a coding harness, but existing options disappointed him on resource use. The timing is therefore tied to his own assessment of local-model capability and harness overhead, rather than to a newly announced model or access change.

## Why it matters
For teams running coding agents on developer machines, CI hosts or remote boxes, the harness is part of the system budget. Lower memory use, if verified in a team's own workload, could leave more capacity for the model, builds or parallel agents. More immediately, Prospero's deadline and process-tree choices address a familiar operational failure: a nominally timed-out command can keep running and interfere with retries, while a long-lived agent can become stuck waiting for a trivial dependency.

The release also separates the harness from the provider. That can make it possible to switch between local and hosted endpoints without choosing the harness based on a bundled model vendor. But this flexibility comes with setup and trust obligations. The operator must supply and configure an endpoint, and should inspect shared configuration includes and credentials. Teams should test the actual integration, permission behavior and cleanup guarantees before assigning unattended tasks.

## Limits and uncertainty
The post is a first-party account by the creator. It gives no benchmark methodology, baseline harness list, hardware configuration or definition of “resources” behind the roughly 100-fold comparison. It does not establish that Prospero is faster, more reliable or safer than alternatives, nor that every listed platform and endpoint combination behaves identically. The material names Community and Professional editions but gives no Professional Edition price. It also does not provide independent validation of desktop control, remote reattachment or process-tree cleanup across operating systems. The configuration safeguards described reduce some risks, but the author explicitly says remote configuration still requires trust.

## Skill to build now
Before moving a coding agent into unattended use, test its timeout and cleanup behavior on a disposable project: run a deliberately slow build that spawns child processes, trigger a timeout, then verify that no build remains and that a retry does not overlap. Prospero makes this failure mode explicit; the same check is useful for any harness your team already uses.

## Reader takeaway
Prospero is worth evaluating not just for its claimed resource footprint, but for its operating model: external model endpoints, headless deployment, bounded waits and process cleanup are treated as core harness features. Treat the “100×” figure as a claim to measure locally, and judge the release by whether its operational controls work in your own tasks.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Anthropic](https://newruntime.com/topics/anthropic.md): Explore the anthropic topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)