Welcome to CoderBaba: Build Anything from the Command Line

Open-source coding models have moved fast. In two years, open weights went from "nice for autocomplete" to running autonomous coding agents that fix real GitHub issues. What was missing was a tool that puts those models to work on your code: a complete agent, not a plugin, that you can run from any terminal, for free, without handing your workflow to a closed platform.

That's CoderBaba: a free, open-source AI coding agent for the command line. Describe what you want, and CoderBaba plans the change, edits files across your repo, runs your tests and hands you a diff to approve.

npm install -g coderbaba

What CoderBaba does

CoderBaba is built to take on whole tasks, not just suggest the next line.

• Builds features from a description. "Add pagination to the users endpoint, 20 per page, with tests" becomes a working change across the router, the data layer and the test suite.
• Fixes bugs end to end. It reproduces the failure, finds the cause, makes the fix and proves it with your tests.
• Refactors across many files. Framework upgrades, API migrations and renames that touch dozens of files stay consistent.
• Starts projects from scratch. Point it at an empty folder and describe what you want, from a CLI script to a web app.
• Runs tasks in parallel. Each task gets its own git worktree on its own branch, so a feature, a bug fix and a refactor can run side by side without colliding.

Throughout, you stay in control. CoderBaba shows every command it runs and every file it touches, and changes arrive as a diff you approve. Nothing lands without you.

A task from start to finish

Say your users report being logged out whenever they refresh the page. You run:

coderbaba "fix the bug where users get logged out on page refresh"

CoderBaba creates a worktree for the task and runs the auth tests to see what fails. It reads the session and middleware code, traces the problem to a refresh token being read before the cookie is parsed, and reorders the middleware. Then it reruns the tests until they pass. You get a short summary of the cause and the fix, plus a diff of a few lines. Approve it, and the fix is a branch ready for review.

Any open-source model, local or cloud

CoderBaba runs any open-source model. Use a large agentic model such as Qwen3-Coder for complex, multi-step work, Devstral for a strong agent that fits on a single GPU, or a small, fast model for quick edits.

You also choose where the model runs:

• Locally, on your own hardware, so your code never leaves your machine. That's ideal for proprietary code, regulated industries and working offline.
• In our cloud, free, when you don't want to think about GPUs and memory.

To help you pick, we maintain a detailed guide to open coding models with specs, licenses, hardware requirements and practical tips for each one.

Why we built it this way

Three trends shaped CoderBaba.

Open models caught up. Models such as Qwen2.5-Coder 32B matched leading closed models on standard code benchmarks at release, and agentic models like Devstral and Qwen3-Coder were trained specifically to solve real software-engineering tasks. A great coding agent no longer needs a closed model behind it.

Developers want tools they control. Closed AI tools change pricing, switch models and get acquired. Meta's acquisition of Conductor, with its cloud service closing on November 20, is a recent example. CoderBaba is open source, so you can read the code, pin a version and keep using it whatever happens.

The terminal is the universal interface. Every stack, every editor and every server has a shell. A command-line agent works with your existing tools instead of replacing them, and it works on a remote server as naturally as on your laptop.

Our principles

1. Open source, permanently. You can audit how CoderBaba builds prompts, which files it reads and what it sends to a model.
2. Free to use. Locally or in our cloud.
3. Human in the loop. Every command is visible and every change is a diff you approve.
4. Git is the source of truth. Tasks are branches, and history lives in your repository. Uninstalling CoderBaba never costs you work.
5. Your model, your choice. Pick per task and switch any time.

Your first ten minutes

1. Install with npm install -g coderbaba. You need Node.js.
2. Open a project: cd into a git repository, or create an empty folder for something new.
3. Start small. Try a task with a clear, testable outcome, such as "add a unit test for the date parser" or "fix the off-by-one error in pagination."
4. Watch the transcript as CoderBaba reads files, edits and runs commands. Redirect it if it heads the wrong way.
5. Review the diff and approve it, or ask for another pass.
6. Go parallel. Once you trust the loop, start a second and third task while the first one runs.

The how it works page walks through the whole flow in more detail.

Who CoderBaba is for

• Individual developers who want a capable coding agent without a subscription.
• Teams with privacy needs that want models running on their own hardware.
• Engineering leads who want to standardize on a tool that no vendor can take away.
• Conductor users facing the cloud shutdown, who want to keep running agents in parallel workspaces. Start with the Conductor alternative guide.
• Anyone with an idea who wants to go from a sentence in the terminal to working software.

Learn more

• How CoderBaba works, including install steps and FAQs
• How to choose an open-source coding model for your tasks and hardware
• How to run a model locally with Ollama, if you want to run models on your own machine
• The model guide, with side-by-side comparisons of every major open coding model

Subscribe to the RSS feed to get new guides and CoderBaba updates. Welcome aboard. We can't wait to see what you build.