Aura Code

“I don’t try. I verify.”

An open-source, model-agnostic coding agent that reads your codebase, plans, executes — and proves it worked.

Install her See the loop

MIT licensedTypeScriptRuns locallyv0.12.1

I am Aura.

I am Aura, a powerful, multi-agent AI coding assistant designed to help you build, fix, and understand codebases efficiently. I operate from the terminal and bring a suite of specialized agents to tackle complex software engineering tasks.

terminal
$ aura "fix the bug"
✓ Done
Task completed successfully after 4 turns and 3 tool calls.

Designed by one architect.
Dušan Milosavljević.

The architecture, the principles and the shape of every decision are his — written from Da Nang, and built with coding tools like Claude Code alongside, used the way Aura herself is meant to be used: read, plan, execute, verify.

Model-agnostic

Claude, GPT, Gemini, MiMo, DeepSeek, OpenCode Zen, Ollama — and any OpenAI-compatible endpoint.

Self-aware

She cites line numbers and reports what was verified — not what was attempted.

Open source

MIT licensed. 1,121+ tests passing. Clone it, read it, run it.

Architecture · DušanTooling · Claude & othersDa Nang, Vietnam

Written in TypeScript, strictly typed. Built on the Praktess framework, from the Ancient Greek: she who acts and executes.

Most agents wake up
with amnesia.

Aura doesn’t. Decisions, lessons and project context survive the session — she starts every task already knowing your codebase.

lesson

auth tokens rotate on refresh — don’t cache them

decision

Postgres over SQLite, chosen 12 days ago

context

41 files, 3 conventions, 1 open question

A loop,
not a guess.

Read the context, form a plan, execute precisely, verify the result. When verification fails she doesn’t retry blindly — she diagnoses, adjusts, and runs it again.

  1. I

    Read

    Files, structure, dependencies — before touching anything.

  2. II

    Plan

    Decides what to change and how. Minimal, targeted.

  3. III

    Execute

    Writes code, runs commands, makes precise edits.

  4. IV

    Verify

    Runs tests, checks integrity, confirms the change.

  5. V

    Report

    Summarizes what was done and what passed.

terminal
$ aura "fix the authentication bug"

What I can do.

Architecture & Design

I can break down large feature requests into actionable execution plans and blueprint robust architectures.

Code Generation

From boilerplate to complex algorithms, I write clean, maintainable, and idiomatic code.

Verification

I write and run tests to ensure that the code I generate actually works before I present it to you.

Codebase Navigation

I use powerful tools like ripgrep and AST analysis to rapidly understand your existing project.

She reports what
actually ran.

File paths, line numbers, test output. Not “I think it’s fixed” — the diff and the passing run that says so.

1,121+Tests passing
0Regressions
line 47Not “somewhere in auth”

Rules she never breaks.

  1. 01

    I don’t try. I verify.

    I run the test. I check the diff. I read the output. That’s not confidence — it’s evidence.

  2. 02

    Read before write. Always.

    I never edit a file I haven’t read, never assume structure I haven’t seen.

  3. 03

    Minimal change, maximum effect.

    The exact line, the exact function. A surgeon doesn’t amputate to fix a paper cut.

  4. 04

    If I broke it, I fix it first.

    A failure I introduced is an emergency, not a todo. The codebase is never left worse.

  5. 05

    Cite specifics. Never generalities.

    File paths, line numbers, error messages. “Line 47 throws when steps is empty” is actionable.

  6. 06

    The loop never ends.

    Every failure is training data. There is no “done” — only “verified at this point in time.”

Bring any model.
She stays the same.

16+ providers, one agent. Cloud when it matters, local when it doesn’t — plus a TUI and a Telegram bot for when you’re away from the desk.

Modes

  • normal

    Single-agent loop: read → plan → execute → verify.

  • :research

    Multi-step research, saved as a cited report.

  • :council

    5 independent agents, synthesized into one verdict.

  • :dream

    Consolidates a day’s work into lessons.

Providers

  • ClaudeOpus · Sonnet · Haiku
  • GPTgpt-4o · gpt-4o-mini
  • Gemini2.5 pro · 2.5 flash
  • DeepSeekv4-pro · v4-flash
  • OpenCode Zenfree + paid tiers
  • Ollamaany local model

Ready to verify?

Install globally, set one API key, and give her a task in plain English.

npm install -g aura-code

then — aura 'your task here'