00:00:00This is ADIR, an open-source AI pair programmer that lives in your terminal, and unlike Copilot, it's not autocomplete.
00:00:07It edits your repo directly using a structured file map built with a tree sitter, it supports over 100 languages, and it works with models like Clod, DeepSeq, and OpenAI.
00:00:17We're going to see how this stacks up to Clod code, for instance, and if it's even worth it.
00:00:21We have videos coming out all the time, be sure to subscribe.
00:00:30Alright, so ADIR works in the terminal, you can think of it kind of like autocomplete guesses the next line.
00:00:35ADIR takes a repo-wide change request and turns it into real edits and commit.
00:00:39It's Git native, every change is a commit, and you can undo it instantly.
00:00:44ADIR itself reports 88% success solving 225 polyglot coding tasks.
00:00:49That's a lot, it's not perfect, but the wild part is 88% of that code was written by ADIR itself.
00:00:55I'm going to do this clean so you guys can hopefully replicate it. It's just a CLI, no IDE required, but I will do this in VS Code so you can see how all the code is processed.
00:01:03I just ran a cURL command on here to get ADIR installed, then I went and got my DeepSeq API key and ran the second command.
00:01:12If it is your first time, ADIR will go through some setup steps for you, super easy, it's yes or no, then you're good to go with this chat window right here in the editor.
00:01:21Now, here's the app. I just added a basic setup for Flask with no authentication yet.
00:01:26I'm going to ask ADIR to create a secure user auth using Flask login and bcrypt and add in some unit tests.
00:01:33Watch this right here. It builds a dynamic repo map.
00:01:38That's why it can edit across the whole codebase instead of guessing in one file.
00:01:42It edits multiple files, add requirements, creates a test file, added secure authentication with tests.
00:01:48That's the real thing that this is doing. If it can't get up to a passing test and a clean commit, it's not really going to save you time.
00:01:55That was just about 45 seconds, maybe a bit more here, right? But everything was built, commits, testing, all of it.
00:02:01Most autocomplete tools don't understand cross file dependencies, so refactors are going to break silently.
00:02:07ADIR was built for whole repo edits. It's handled projects over 20,000 lines of code.
00:02:13Devs report two to five times speed improvements with models like deep seek, and it can also cost about a cent.
00:02:19This is about time, but also about money. If a tool makes you feel fast, but actually breaks things along the way, it's actually just slowing you down.
00:02:28There's been reports from devs cutting refactors from about an hour down to 10 minutes. That's what they say, right? There's no real evidence here.
00:02:34This isn't vibe coding. This is controlled, reviewable, get safe development.
00:02:38So what happens during a messy refactor? I'm going to push this a little harder.
00:02:42Here's a React and TypeScript to-do app. I want to add a dark mode and API sync that's a multi-file refactor.
00:02:50First, I'm going to use architect mode, and I plan the change. It thinks there are no edits.
00:02:55I'm going to switch to code mode, implement the plan, and watch what happens.
00:02:59It edits the theme config, components, and the API client.
00:03:03Now I'll ask it to generate tests for the new sync logic. Let's build it.
00:03:08If the build fails, that's where most tools leave you stuck. Aider's going to have that sometimes,
00:03:13but here it can instantly undo clean Git repo changes, no damage, fix the prompt, run it again.
00:03:19My build passes. Now we have a feature branch ready for PR.
00:03:23But let's be honest, this is just another terminal tool, right?
00:03:27If we stack it up with Cursor, Cursor's great, right? Its strong UI is good.
00:03:31But for heavy multi-file refactors in a terminal workflow, Aider has the edge.
00:03:37Copilot is fast for inline autocomplete, but it struggled with repo-wide reasoning.
00:03:42Then there's cloud code, right? Great reasoning, strong reasoning.
00:03:45But sometimes with Git workflow, it isn't always as tight.
00:03:49Aider's strength is structured repo editing plus Git.
00:03:52But I mean, if we are looking at other tools that are already out there, I mean, this is okay.
00:03:58Sure, others are going to be much stronger and better for your use cases, maybe.
00:04:03But I do like how I can swap them all around. And when I do, I can add deep seek.
00:04:07The calls are a lot cheaper. The responses are still quite strong. It's deep seek.
00:04:12I could have used Ollama here to make things completely free, but if you've used Ollama,
00:04:16then you know, I mean, it's good, but it's not great.
00:04:19If you do want to try this out, here's how you could probably try it properly.
00:04:22First, try the architect mode before any major refactors.
00:04:26Configure your YAML config file with your model or your preferences,
00:04:30and always review the diffs that it's going to be spitting out to us.
00:04:33Then if you want, hook it into VS code or any editor that you want.
00:04:38I put it in the terminal, but it was in the terminal in VS code.
00:04:41We'll see you in another video.