# Prospero Coding Harness: Local Models, Remote Agents and Pricing

Chris Wellons' native coding harness separates the agent interface from the model, with a free community build and a paid remote-engine bundle.

By Paras Madan | 2026-10-31


Chris Wellons introduced Prospero on October 31, a native AI coding harness that connects local or hosted models to development tools. It supports subagents and remote execution, while leaving the model choice to the user. The professional edition costs $25 or more; a community version is free.

The announcement follows a store listing published on October 29 and updated on October 30. This is an early tool, with version 0.9.0 offered for download, rather than a claim that every part of the product is finished.

## What Prospero does

A coding harness gives a model access to tools and carries their results back into the conversation. That loop lets an agent read files, make changes and run commands without asking the user to copy every output into chat.

Prospero adds a native C++ interface built with Dear ImGui. Its [repository](https://github.com/skeeto/prospero) documents compatibility with OpenAI-style Chat Completions endpoints and Anthropic's Messages API. That includes local servers such as llama.cpp, as well as compatible remote providers.

No provider comes configured. You supply the endpoint and access yourself, and a Claude consumer subscription is not a substitute for API access. **The harness purchase does not buy model usage.** Local inference still needs suitable hardware; hosted inference still has its own pricing.

## Prospero pricing and availability

The [official store](https://skeeto.itch.io/prospero) offers the Professional Edition for a minimum of $25, with version 0.9.0 downloads for macOS on Apple Silicon and Windows on x86-64. The store labels the project as in development.

The free community build and public-domain source are available through GitHub. The paid bundle includes embedded remote engines for supported architectures. That packaging difference matters for people who want to run agents on machines other than the one displaying the interface.

Prospero itself supports Windows, macOS and Linux, but that does not mean the paid store currently lists a desktop installer for every combination. Check the available build for the machine that will run the interface, separately from the remote engine you need.

## Remote agents without a remote desktop interface

Wellons' [launch explanation](https://nullprogram.com/blog/2026/10/31/) describes deploying a headless engine over SSH or WSL. The engine runs where the files and commands live, while the graphical interface remains on the user's desktop. With additional setup, execution can continue after the interface disconnects.

This is useful when a project belongs on a Linux development machine but the person supervising it works elsewhere. It also makes two resource questions easier to separate: what the interface consumes, and what the model or development workload consumes.

Wellons claims a large resource advantage over other harnesses. That is the developer's claim, not an independently verified benchmark here. A smaller interface also does not shrink the model's weights or the memory required by a compiler.

## Tool permissions are not a sandbox

Prospero supports desktop control with a vision-capable model, alongside file and shell tools. That expands what an agent can test, but also expands what it can affect.

The author explicitly treats permissions as separate from security isolation. **An approval prompt is not an operating-system boundary.** If the agent should only see a disposable working copy, that restriction needs to exist in the environment where it runs.

There are other deliberate tradeoffs. Prospero relies on external editors and terminals instead of bundling its own, and its timeout behavior is designed to keep unattended work from waiting indefinitely. These choices suit a developer assembling a workflow more than someone expecting one subscription to supply the model, editor and execution environment.

For local-model users, the useful evaluation is whether the same task completes reliably with less overhead and a manageable permission setup. Keeping the model fixed during that comparison would reveal what the harness contributes, rather than crediting the interface for a stronger model underneath it.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders explain product launches and engineering choices in content developers can understand and share.

[Book a 30-minute call](https://cal.com/paras-madan/30min).