Aider

What it is

Aider is one of the original open-source AI coding tools — it predates most of this board and established conventions the newer harnesses inherited without always crediting. Apache-2.0, Python, around 48,000 GitHub stars.

The framing is pair programming rather than delegation. You sit with it, you describe a change, it makes the change, you look at the change. It was never designed for the walk-away-and-come-back-later pattern that dominates 2026 harness discussion, and it does not pretend otherwise.

How the loop works

Two ideas carry most of the weight. The first is the repo map: Aider indexes the project structure and feeds the model a condensed picture of it, so the model has a sense of where things live without reading the entire tree. That is how it coordinates edits across several files at once rather than fumbling one file at a time.

The second is git. Every edit lands as its own commit, written by Aider with a message describing what it did. This is the design decision that still distinguishes it. Your history becomes a step-by-step record of the session, a bad change is one git revert away, and reviewing the agent's work is the review workflow you already have rather than a new one.

Model choice is yours — bring your own key, any provider.

Isolation and permissions

There is no sandbox, and Aider's answer is that it does not need one in the usual sense: the commit granularity is the containment. Nothing accumulates into an unreviewable pile of changes, because each step is separately recorded and separately revertible. That protects your repository well. It does not protect anything outside it, so the normal precautions apply to commands you let it run.

Who it is for

Developers who want a strong assistant for a focused hour and a git log they would not be embarrassed to push. It is particularly good in codebases where review discipline matters, because the output arrives pre-segmented into reviewable units.

Aider also publishes an open polyglot coding benchmark — 225 Exercism exercises across six languages — which is one of the more transparent public comparisons available for model coding ability, and worth knowing about independently of whether you use the tool.

Limits

It is built for supervised interactive sessions, and the wider community's harness reviews have deliberately excluded it from long-unattended-run comparisons on those grounds. That is a fair exclusion rather than a criticism: it is being left out of a contest it was not designed to enter. If your requirement is an agent that runs for hours on its own, look elsewhere.

Release cadence has also slowed through 2026. The project works, but it is no longer setting the pace it once did, and features that newer harnesses treat as standard — background sessions, sub-agents, MCP tooling — are not its focus.

Alternatives on this board

• Claude Code — the supervised-to-autonomous step up, with real permissions.
• OpenCode — provider-agnostic terminal agent with a more current feature set.
• Editor agents — the same tight feedback loop, inside the editor.

Sources

• aider.chat
• Aider-AI/aider on GitHub
• Aider polyglot coding benchmark