Aurora

A micro terminal coding agent, small enough to read and complete enough to use every day. Named after my beautiful two-year-old daughter Aurora, always with her hair bow.

Aurora is a coding agent that lives in the terminal. You talk to it in plain text; it reads and edits files, runs commands and searches the web through a tool loop, always asking before it changes anything. Drop into a persistent bash shell or its built-in editor without leaving the session. The same agent also runs as a small web app, so you can use it from a browser or a phone. I built it from scratch in Python to understand how agents like Claude Code work, and it grew into a daily tool.

Version 1.3.0 · Source on GitHub

Aurora at startup: local model ready with its context size, extensions loaded, and the offer to run the session bootstrap prompt

Features

Aurora's secret-detection challenge, catching an AWS key, GitHub token, Bearer token and UUID in a file before it reaches the model

Privacy and security

Every prompt and every tool output (even read-only ones like read_file/grep) is scanned for API keys, tokens and credentials before it reaches the model or the session log. A match lets you redact it, keep it, allow it going forward (remembered, never flagged again), or stop.
API keys never touch disk in plaintext. They live in the OS keyring or encrypted storage.

Approvals and edits

Every file write, edit or shell command shows a diff preview and asks first; trusted patterns can be allowlisted.
Instead of allowing or denying, pick "Comment" and type a note. Aurora folds it into the same request and redoes the change, so you skip the deny, re-explain, re-approve round trip.
Pick "Explain" and Aurora describes what the change will do before asking again, so you never approve blind.
Aurora takes a shadow git snapshot before every approved change, so /rewind can undo any step, even the rewind itself.
Trust the task? /auto-approve on skips the prompt until you turn it off again.
/plan on lets Aurora read, search and fetch but refuses every write, edit and command, even ones you allowed, so it writes the steps down instead of doing them.
/try <task> runs the task in a throwaway git worktree. You see the diff and choose apply, discard or keep, and your real tree stays untouched until you apply.
/budget <usd> caps spending for the session or the day. Aurora warns at 80% and stops calling paid models at 100%.
/commit drafts the commit message from your diff and commits once you approve it, after scanning the diff for secrets.

Aurora explaining what a proposed shell command will do before asking for approval again

Any model

OpenRouter, or a self-hosted llama.cpp or Ollama server, switched from the same arrow-key /model menu. The current model is marked, and if no key is stored yet it asks once and remembers.
Adding one is just pasting a URL: /model add https://openrouter.ai/<org>/<model> validates it against the OpenRouter catalog, fetches its context size, pricing and description, asks for your key if it's missing, and switches to it, with no config editing. /model remove drops one just as easily.

The /model picker: Claude, Kimi and GPT via OpenRouter with live prices, the local llama.cpp model with its measured latency, and Ollama models

Daily terminal use

Full-screen TUI: scrollable chat with streaming markdown, timed collapsible thinking blocks, mouse support and drag-to-copy, with a classic REPL fallback for plain terminals.
Resume past conversations, export them as markdown, /compact long ones into a summary, or /remember what mattered into a persistent cross-session memory.
Reusable skills (plain-Markdown SKILL.md files work too), plus a per-project bootstrap prompt so the agent starts every session already knowing the codebase.
Keep typing while a turn is running, and your next message is ready when the answer lands.
Prompt caching and parallel read-only tool calls, so a multi-step task doesn't burn tokens or time.
Extend it with a plain .py file: an extension API with an MCP client and a lint checker bundled.
Drop into a persistent bash shell with folder autocomplete and real cd tracking, without leaving the session.
A built-in /nano editor for quick edits, so you don't shell out to a real editor.
Built-in web_search/web_fetch tools when the answer isn't in your codebase.
/cost shows a per-model token and $ breakdown, cache hits included.

Persistent bash mode in Aurora: cd tracked across commands, ls output linkified

Aurora using its web_search and web_fetch tools to answer a live weather question

Aurora on the web

aurora --web serves the same agent as a chat page you can open in any browser, on a desktop or a phone. It uses the same tools, models and approvals as the terminal, and has a few extras of its own.

Aurora's web page: a sidebar with chats, Notes, Files and search, and a finished turn where the local model added a function to a file

A sidebar of past chats with search, rename and drag-and-drop folders. Any chat can be downloaded as one HTML file, either just the prompts and answers or everything including thinking and tool output.
Send several messages without waiting: they run one at a time, in order. Replies have Regenerate, Edit, Save as note and Read aloud.
Markdown notes in Personal, Work and Skills tabs, with tickable tasks. Type # in the message box to send a note to Aurora as context.
Attach or drop documents, pictures, music and video. Files has Input and Outputs tabs with a preview for common file types, and pictures Aurora makes show inline with a Download button.
Tap the model name to switch, wake a sleeping local model, or edit the model list and the config as YAML. Each change is validated first, the last few versions are kept, and your keys are never shown.
Voice input, a plan mode switch, and a context meter that shows how full the session is.
Light and dark themes, an installable home-screen page, and it works fully on a phone.
The page needs a bearer token and binds to loopback by default. Put it behind a TLS proxy, because the page can run tools.

The same Aurora chat on a phone, with the reply actions as icon buttons

Install

Source on GitHub · Read the installer first

curl -fsSL https://raw.githubusercontent.com/ricardopsantos/Aurora/main/bootstrap-install.sh | bash

It runs on macOS or Linux, degrades gracefully when the local model server is unreachable, and the whole spec (requirements, build plan and test plan) ships in the repo.