Aider, three weeks in: a pair programmer that respects git

I used it for real tickets on a Node API and a pile of Terraform. It earned a permanent place in my terminal, with some rules attached.

Sample post. This shows the review format. The testing notes are written as an example; treat the observations and any figures as illustrative, not measured results.

Use it when the work has to land as clean, reviewable commits. Skip it if you want something to drive the whole task for you.

Verdict: Use it
Cost: Free toolYou pay the model provider per token; local models cost power.
Licence: Apache-2.0Open source, runs on your machine.
Tested: 3 weeksNode/TypeScript API, Terraform and Ansible.

What I tested, and how

I ran Aider as my main assistant for three weeks of ordinary work. No toy projects: every task came from a real backlog, and every change went through the same review I would give a colleague's pull request.

01Feature work on a Node/TypeScript API. New endpoints, validation, and the tests that go with them.
02Refactors across many files. Renaming a domain concept and moving modules without breaking imports.
03Infrastructure as code. Terraform modules and Ansible roles for a small homelab cluster.
04Two model set-ups. A hosted frontier model for hard tasks, a local model through Ollama for small edits.

My rule throughout: I accept nothing I have not read. If a tool makes reading its output harder, it fails, however clever the output is.

What it gets right

It works in git, not around it. Every change Aider makes is committed with a descriptive message, so the history reads like a careful colleague's. When it goes wrong, /undo drops the last commit and I am back where I started. That one design decision is why I trust it more than tools that leave a heap of uncommitted edits.

You choose what it can see. You add files to the chat explicitly, and it builds a map of the rest of the repository for context. On the API this kept edits focused. It also kept my token spend predictable, which matters once you read the monthly market post.

Any model, including local ones. Switching between a hosted model and a local one is a flag. Small edits on the local model were good enough that I stopped paying for them.

The best thing I can say about it: reviewing its work felt like reviewing a pull request, not like auditing a magic trick.

How it compares

I ran the same kinds of task through three other open-source tools I already had installed. This is how they differ in daily use, not a benchmark.

Open-source coding assistants, side by side

Licences as stated by each project · my verdicts

Tool	Verdict	Licence	Where it runs	What happens to your changes	Local models	Best for
AiderThis review	Use it	Apache-2.0	Terminal, any editor	Committed to git per change, with a message; one-step undo	Yes, via Ollama and others	Focused changes you will review line by line
ContinueEditor extension	Use it	Apache-2.0	VS Code, JetBrains	Edits applied in the editor; you commit	Yes	Inline help and chat while you drive
ClineEditor agent	Worth watching	Apache-2.0	VS Code	Agent edits files and runs commands with approval; checkpoints to roll back	Yes, quality varies	Longer tasks you are willing to supervise
OpenHandsSandboxed agent	Worth watching	MIT (core)	Docker sandbox, web UI or CLI	Works in an isolated workspace; you pull the result	Yes, needs a strong model	Whole tasks you can describe and check later

Scroll the table sideways

Where it falls short

It needs you to drive. Aider does what you ask, in the files you give it. If you want a tool that plans a whole feature and runs the tests on its own, look at the agents in the table instead.
Terraform needs a firm hand. It wrote plausible modules that would not have passed terraform plan without fixes. Run the plan every time; do not trust a confident commit message.
Small local models drift on large edits. Fine for a function, unreliable for a cross-file refactor. Use the hosted model for those and accept the cost.

The verdict

Aider is the tool I reach for when the change matters and I want to understand every line of it. It is honest about what it did, easy to undo, and cheap to run. It is not an autonomous engineer, and it does not pretend to be.

Use it Keep it on the hosted model for refactors, the local one for small edits, and read every commit.