# Harneloop


> Helps an AI agent learn from what went wrong: inspect real results, trace the mistake and keep only harness changes a test supports.


Canonical page: https://toolindex.org/tools/harneloop/


Type: CLI
Category: Coding
Version: 0.0.2
License: Apache-2.0


Source: https://github.com/OpenRootdev/Harneloop
Distribution: https://pypi.org/project/harneloop/0.0.2/
Scope: The harneloop command-line tool as a Python package. The repository also holds docs, examples, tests and translations that are not needed to use it.


## Installation


### 1. Install the command

Install it as an isolated tool with uv, then run the doctor check. You can also use python -m pip install harneloop in your own Python 3.11+ environment.

Harneloop on PyPI: https://pypi.org/project/harneloop/

```

uv tool install harneloop
harneloop doctor

```


### 2. Give your agent the repository

Harneloop is meant to be driven by an agent with file and terminal access. Ask it to read the README and docs/agent-onboarding.md, then use Harneloop to create a harness unit for the task you want it to get better at.

Agent quick start prompt: https://github.com/OpenRootdev/Harneloop#agent-quick-start

```

harneloop onboard --format json

```


### 3. Or start it yourself

harneloop setup opens a guided menu that creates and registers a harness unit. After that you can check any unit by name from any folder.

```

harneloop setup
harneloop units list

```


## Manual setup

- Python 3.11 or newer and git. uv is recommended for an isolated install; if uv says its tool folder is not on PATH, run uv tool update-shell and open a new terminal.

- You bring the task and the environment it runs in. Harneloop does not discover endpoints, tools, screenshots or output paths for you; the agent maps them into the harness unit.


## Additional notes

- Public alpha. The core loop works; the first published case study is Blender scene construction with ViperMesh.

- It changes the harness around the model (instructions, tools, context and checks), not the model itself. A proposed change becomes the working version only after a test supports it.


## Contents

- skills: 1

- tools: harneloop CLI, Python 3.11+, git


## Review scope

setup-reviewed; reviewed 2026-11-01.

Windows: installed harneloop 0.0.2 from PyPI in an isolated uv environment and harneloop doctor passed (Python, PyYAML, git, writable folder). README install and agent quick start reviewed. A full harness-unit run was not repeated for this listing.

Evidence source: https://github.com/OpenRootdev/Harneloop/releases/tag/v0.0.2
