CoderBaba: free, open-source AI coding agent for your terminal

CoderBaba is a free, open-source AI coding agent that runs in your terminal. Describe what you want in plain English, and CoderBaba reads your project, plans the work, edits files, runs commands, starts your dev server, checks the page in a real browser, and fixes the errors it finds. It asks before every change, and /undo reverts any turn.

npx coderbaba

No sign-in, API key, credit card, or config file needed. You get 1,000 free tasks per device on the Fast model. Log in with your email (also free) to unlock the Smart model.

Website · npm · YouTube · X (Twitter) · MIT license

---

Contents

• Who CoderBaba is for
• What CoderBaba can do
• CoderBaba Security
• Supported stacks
• Models
• Safety and permissions
• Privacy
• Commands
• Configuration
• FAQ
• Packages and development

Who CoderBaba is for

• Beginners and students who want to build a real app without learning every tool first. CoderBaba explains each step in plain language before it does it.
• Developers who want a terminal coding agent that runs the app, checks it, and fixes it, instead of only writing code.
• Anyone who wants a free AI coding assistant built on open models, with no subscription and no API key.

CoderBaba works on macOS, Linux and Windows, and needs Node.js 20 or newer.

What CoderBaba can do

Build apps from a plain-English description

Ask for "a notes app with login" or "a Django API with a React frontend" and CoderBaba scaffolds the project, installs dependencies, writes the code, and starts it. For tasks with more than a few steps it writes a todo list first and updates it as it works, so you can follow the plan.

Read, search and edit your code

CoderBaba lists folders, reads files and searches your code with grep before it changes anything. It edits files with targeted changes rather than rewriting whole files, and it never invents the contents of files it hasn't read.

Run commands, dev servers and databases

• Runs shell commands such as npm install, pip install or git status, and reads their output.
• Starts long-running processes (dev servers, API servers, watchers, databases) in the background, detects their port and URL, and restarts them after changes.
• Notices when a port is already in use and tells you.
• /ps, /stop and /restart let you manage those processes yourself.

Check its own work automatically

After every edit, CoderBaba runs your project's own checks and fixes what they report, up to two rounds:

• TypeScript type checks and ESLint.
• Django's manage.py check.
• Dependency audits (npm audit, pnpm audit, yarn audit or pip-audit) when you change dependency files. These warn you about known vulnerable packages without blocking your work.

Checks are found per app folder, so a repo with client/ and server/ gets the right checks for each.

Test pages in a real browser

When a web page is running, CoderBaba opens it in headless Chrome or Edge and catches uncaught errors, console errors, crashes and failed network requests, which curl can't see. It can take a screenshot when you want to see the page. If neither browser is installed, it asks before downloading Chromium.

Read documentation

CoderBaba reads documentation pages with read_url when a command fails on changed flags or it needs current package versions. Well-known docs sites open without asking, including:

• nextjs.org, react.dev, vite.dev and tailwindcss.com
• docs.djangoproject.com and docs.python.org
• developer.mozilla.org, nodejs.org, expressjs.com and mongoosejs.com
• npm, PyPI and GitHub

Other sites ask once per host, and private or local addresses are blocked.

Handle secrets safely

CoderBaba never asks for passwords or API keys in chat. When your app needs a secret, it prompts you for it directly and saves it to .env, so the model never sees the value. It keeps .env out of git, writes a .env.example with names only, and redacts secret values from command output.

Remember your project

• AGENTS.md project notes: CoderBaba reads AGENTS.md at the start of every turn and keeps its "How to run" section current. /init writes a first version for you.
• Chat history: chats are saved on your machine, and /history lists recent ones.
• Long chats: these are summarized automatically when they get too big for the model, and /compact does it on demand.
• /export: saves a chat as Markdown.

Undo anything

CoderBaba saves a checkpoint of every file it touches. /undo reverts the last turn's file changes, even if your project doesn't use git.

CoderBaba Security

CoderBaba includes an AI security scanner that reviews your whole project the way a security engineer would, then helps you fix what it finds.

Command	What it does
/security	Deep scan. Maps your app first (routes, how login works, where data lives, where untrusted input enters), then reviews each app folder against that map and writes SECURITY_REPORT.md.
/security validate [id]	Tries to reproduce findings against your app running on localhost, and marks each one confirmed or a false positive.
/security fix <id>	Fixes one finding, runs your checks, and verifies the fix.
/security diff [base]	Reviews your uncommitted changes, or your branch against a base branch, before you push.
/security status	Lists findings and their status.

It looks for the issues that cause most real breaches in small apps:

• Missing authentication or authorization checks, including records loaded by id that don't belong to the current user (IDOR).
• SQL and NoSQL injection, command injection and eval of user input.
• Cross-site scripting (XSS) through dangerouslySetInnerHTML, innerHTML, v-html or Django's |safe.
• Passwords stored or logged in plain text.
• Hard-coded secrets, unsafe CORS settings, and Django DEBUG left on.
• Known vulnerable dependencies.

Findings are stored in .coderbaba/security/, which is kept out of git. Secret values are redacted from the report.

Secure by default: even without /security, CoderBaba writes code with parameterized queries, hashed passwords (bcrypt, argon2 or Django auth), auth checks on routes that change data, and request validation (zod, express-validator or DRF serializers).

Supported stacks

CoderBaba works with any project it can read and run. It has built-in, tested setup steps for:

• Next.js (TypeScript, Tailwind, App Router).
• MERN: MongoDB, Express, React (Vite) and Node, with an API proxy so there are no CORS errors.
• Django + React, and Django on its own.
• Python projects with automatic virtual environments, using uv when it's installed and venv otherwise. It never runs pip install outside the virtual environment.
• Node.js projects using npm, pnpm or yarn.

On Windows it runs commands in cmd.exe and uses Windows forms of commands. On macOS and Linux it uses sh.

Models

CoderBaba uses open models only, served through CoderBaba's free cloud:

Model	Runs	Who can use it
Fast	Cohere North Mini Code	Everyone, no sign-in
Smart	Qwen3-Coder-Next	Free after email login

Pick a model with /model. Contributors can also bring their own OpenRouter key (see Configuration).

Safety and permissions

• Asks before acting. Before every file edit or command, CoderBaba shows what it wants to do and asks:

  CoderBaba wants to run npm install. Allow? (y / always / n)
  

• Safe read-only commands run without asking. Examples are git status, git diff, version checks, npm audit and curl to your own localhost. /allow shows the full list, and you can add your own.

• Stays in your project folder. It won't read or write outside the project unless you allow it.

• No interactive commands. stdin is closed, so commands can't hang waiting for input.

• Undo. /undo reverts the last turn's file changes.

• Project config needs approval. If a project's .coderbaba/allow.json adds allowed commands or checks, CoderBaba asks you once, and again whenever the file changes. A cloned repo can't silently grant itself permissions.

Privacy

• The CLI talks only to the CoderBaba API at https://coderbaba.si. The model API key lives on that server, never on your machine.
• Requests are sent with OpenRouter's data_collection: "deny", so they only go to model providers whose policy is not to collect or train on your prompts.
• Secret values from .env files are redacted from everything CoderBaba's tools return, including file reads and command output, so the model doesn't see them.
• Your login token is stored in ~/.coderbaba/auth.json with owner-only permissions (mode 600).

Commands

Command
/help	show commands
/new	start a new chat
/history	list recent chats
/usage	show your free usage (also at coderbaba.si/usage)
/model	pick Fast or Smart
/undo	revert the last turn's file changes
/compact	summarize the chat so far to free up room
/init	write AGENTS.md, notes that help CoderBaba understand this project
/security	scan this project for security issues (see CoderBaba Security)
/todos	show this chat's todo list
/export	save this chat as a Markdown file in the project
/ps, /stop, /restart	list, stop or restart background processes such as dev servers
/allow	show which commands run without asking, and which checks run
/login, /logout	log in or out on this machine (free)
/exit	quit
!command	run a command yourself, such as !npm test; CoderBaba sees the output in your next message

While CoderBaba works you can keep typing: Enter queues your next message and Esc clears the queue.

Configuration

CoderBaba works with no configuration. To customize it:

Project notes

Put an AGENTS.md at your project root with how to run the app, which checks matter, and your conventions. CoderBaba reads up to 8,000 characters of it every turn.

Checks and allowed commands

Add a checks list to .coderbaba/allow.json in your project, or to ~/.coderbaba/config.json for all projects:

{
  "allow": ["npm test"],
  "checks": [
    { "command": "npm run typecheck", "cwd": "web", "files": "web/src/**/*.ts*" },
    { "command": "python manage.py check", "cwd": "api" }
  ]
}

Configured checks replace the detected ones. files limits a check to matching edits, and cwd is the folder it runs in. Set "checks": false to turn checks off.

Environment variables

CODERBABA_API_URL=http://localhost:3000 npx coderbaba   # use your own CoderBaba server
OPENROUTER_API_KEY=sk-or-... npx coderbaba              # call OpenRouter directly: no login, no usage tracking
CODERBABA_HOME=~/somewhere-else npx coderbaba           # move local state (default ~/.coderbaba)

Local state lives in ~/.coderbaba: auth.json, sessions/ and checkpoints/. Page screenshots go to .coderbaba/screenshots/ in your project, which has its own .gitignore.

FAQ

What is CoderBaba?

CoderBaba is a free, open-source AI coding agent for the terminal. You describe a task in plain English, and it edits your code, runs commands, starts your app, tests it in a browser, and fixes errors, asking your permission before each change.

Is CoderBaba free?

Yes. The CLI is MIT-licensed, and the hosted models are free: 1,000 tasks per device with no sign-in, and more, plus the Smart model, with a free email login. There is no paid plan.

Do I need an API key or an account?

No. Run npx coderbaba and start typing. An account is optional and only unlocks the Smart model. If you prefer, you can use your own OpenRouter key instead.

Which AI models does CoderBaba use?

Open models only: Cohere North Mini Code for Fast and Qwen3-Coder-Next for Smart, served through OpenRouter with data collection denied.

Is CoderBaba good for beginners?

Yes, that's who it's designed for. It explains each step in plain language, asks before every change, sets up virtual environments and dev servers for you, handles secrets without showing them in chat, and /undo reverts any mistake.

Can CoderBaba find security vulnerabilities in my code?

Yes. /security runs a deep scan and writes SECURITY_REPORT.md. /security validate reproduces findings against your running app, /security fix <id> fixes one, and /security diff reviews changes before you push.

Does CoderBaba work on Windows?

Yes. CoderBaba runs on Windows, macOS and Linux with Node.js 20 or newer.

Is my code used to train AI models?

No. Requests use OpenRouter's data_collection: "deny" setting, which only routes to providers whose policy is not to collect or train on prompts, and secret values from .env are redacted before anything reaches the model.

How is CoderBaba different from an AI chat assistant?

A chat assistant gives you code to copy. CoderBaba is an agent: it works inside your project, makes the changes, runs the app, checks the result with your type checker, linter and a real browser, and fixes problems before it says it's done.

Can I use CoderBaba with my own server or model key?

Yes. Set CODERBABA_API_URL to point at your own CoderBaba server, or OPENROUTER_API_KEY to call OpenRouter directly.

Packages and development

Package	npm
packages/cli	coderbaba	Ink terminal UI and slash commands
packages/core	@coderbabasi/core	UI-free agent loop, tools, permissions, checks, security scans, /undo, API and login client, sessions
packages/shared	@coderbabasi/shared	API types, model allowlist, error codes

core emits events and asks for permission through an injected askPermission() callback, so other front ends can reuse it.

Development requires Node.js 20+ and pnpm 9:

pnpm install
pnpm build
pnpm test
node packages/cli/dist/bin.js

See CONTRIBUTING.md to contribute and SECURITY.md to report a vulnerability. Releases go through Changesets; see .changeset/README.md.

License

MIT. Made by CoderBaba.SI. Follow along on YouTube and X.