Explore AI
Harnesses
The model is the engine. The harness is the car — and it matters more than the engine does.
What a harness actually is
A model on its own produces text. A harness is the program that puts that text to work: it reads your files, runs commands, looks at what happened, and tries again. It is the difference between asking someone for advice and handing them the keyboard.
This matters more than the model choice, and it is the least discussed part of the whole field. The same model inside a good harness will outperform the same model in a chat box by a wide margin — because the harness supplies tools, memory of the task, and the ability to check its own work.
This category churns faster than any other
Every number on this page is a real measurement from the GitHub API, including the last-commit date — and that date is the column to read. At the time of checking:
- Roo Code is archived. A popular fork of Cline, ended.
- Sourcegraph Cody has gone — the repository returns a 404.
- Aider's last commit was months ago. Not dead, but no longer moving fast.
- The Stable Diffusion web UI is roughly six months stale, superseded by ComfyUI.
None of that makes those tools bad. It means a harness is a commitment to something that may not exist in two years, and that the abundance of choice is partly an illusion — most of these will not be here.
Prefer a harness that lets you bring your own model, and that keeps your work in files you control — your repository, your git history. If the tool disappears, you keep everything that mattered. That single criterion rules out very little here and protects you from all of it.
If you only try one
Install opencode or Gemini CLI — both are permissively licensed, both are actively developed, and Gemini CLI has a free tier that makes experimenting cost nothing. Point it at a small task in a repository you do not care about, and watch what it does before you trust it with anything.