austere-axi

mcp
Guvenlik Denetimi
Basarisiz
Health Uyari
  • License — License: MIT
  • Description — Repository has a description
  • Active repo — Last push 0 days ago
  • Low visibility — Only 8 GitHub stars
Code Basarisiz
  • exec() — Shell command execution in action/lib.ts
  • process.env — Environment variable access in action/lib.ts
  • process.env — Environment variable access in action/report.ts
  • spawnSync — Synchronous process spawning in action/run.ts
  • fs.rmSync — Destructive file system operation in action/run.ts
  • process.env — Environment variable access in action/run.ts
  • fs.rmSync — Destructive file system operation in bench/lib/paths.ts
  • exec() — Shell command execution in bench/lib/prompts.ts
  • process.env — Environment variable access in bench/lib/prompts.ts
Permissions Gecti
  • Permissions — No dangerous permissions requested

Bu listing icin henuz AI raporu yok.

SUMMARY

Coding agents write review pages in OpenUI Lang (a few hundred tokens, not a whole HTML page). You annotate them in the browser; notes go back by statement name and the agent patches only those. Local, MIT.

README.md

austere-axi

npm version
CI
license

Review what your AI coding agent proposes on a page you can click, instead of in a wall of chat text.

An agent writes a page, the reviewer clicks a table cell and a diagram box and sends two notes, and the page updates in place with the changed parts flashing

A real review, about 30 seconds: MP4 · WebM.

See it without installing anything: example pages · try it in your browser · website

The problem

You ask an AI coding agent (Claude Code, Codex, Cursor and the like) for a plan, a comparison, or a summary of what it changed. The answer comes back as a long chat message.

Reviewing that message is slow:

  • You have to describe which part you mean ("the second option in the rollout table") before you can say what is wrong with it.
  • The agent guesses what you pointed at, then often rewrites the whole answer, so you read all of it again to find what changed.
  • A diagram or a side-by-side comparison would make the decision easy, and chat gives you paragraphs.

What austere-axi does

The agent puts its answer on a page: sections, tables, diagrams, charts, code changes, and a form for the decision it needs from you.

You click the exact thing you want changed and write a note, the way you would comment on a shared document. The agent changes only that part, and the page updates while you watch.

How you use it

  1. Ask your agent for a page. Say "show me this plan as a page", or type /austere followed by what you want to see. The page opens in your browser.
  2. Click what you want changed and write a note. Anything on the page works: a section, one cell of a table, one box in a diagram, one line of a code change, or a sentence you select. Where the page offers choices, pick one or fill in the form.
  3. Press Send. The agent gets each note attached to the part you clicked and changes only those parts. The page updates in place and highlights what changed, so you don't have to hunt for it.

Repeat until it is right. When you are done, press Send & End and carry on in the chat.

What people review with it

Example
A plan, before any work starts Moving a database, with a diagram, rollout options and a decision form
A design, with alternatives to choose from API design review
A summary of what the agent changed Pull request summary
A report on something that went wrong Outage postmortem
A plan for the team Sprint plan

What else it does

  • Bring in a teammate. Ask your agent to publish the page and invite someone by name. Each person gets their own link, and their notes appear under their name. You see their notes first and decide which ones go to the agent. More on sharing.
  • Review from your phone. The review page has a phone layout. This needs Tailscale on your computer and phone. How it works.
  • Send someone a copy. "Download standalone HTML" in the page's ⋯ menu saves the page as one file that opens in any browser, offline.
  • Nothing gets lost. Every note is saved the moment you send it. If the agent is busy or your computer restarts, the notes are waiting when the agent comes back.
  • Find old reviews. Turn on the hub for a home page that lists every page and every note, with search. More on the hub.

Get started

You need two things:

  • An AI coding agent, such as Claude Code, Codex or Cursor.
  • Node.js version 22.18 or later. It is free, and you install it once.

Then:

  1. Run this once in a terminal:

    npx skills add MandalorianBatman/austere-axi --skill austere-setup
    
  2. Tell your agent: "set up austere-axi". It checks your computer, installs the rest, and opens a test page so you can see it work.

After that, ask for a page whenever an answer would be easier to review as one.

Using Claude Code? You can install it as a plugin instead:

claude plugin marketplace add MandalorianBatman/austere-axi
claude plugin install austere@austere-axi

Claude Desktop, Cursor, Codex and manual setup are covered in the install guide.

Why it is quick

Other tools have the agent hand-write a whole web page, which is slow, costs more, and looks different every time.

With austere-axi the agent writes a short outline that names what goes on the page, and austere-axi draws it. Three things follow:

  • Pages arrive sooner and cost less. In our benchmark, an agent hand-writing the same pages as full web pages took more than twice as long and cost more than twice as much. The benchmark also lists the cases where the other approach did better.
  • Every page looks the same. One clean, dark design, so you learn where things are once.
  • Changes are small. The agent edits the one part you pointed at instead of rebuilding the page.

Private by default

  • It runs on your own computer. There is no account and no sign-up.
  • Your pages and notes stay on your computer unless you choose to publish a page.
  • A published page is invite-only unless you ask for an open link, and it expires 30 days after you last published it.
  • It is free and open source (MIT license).

The details are in the security guide.

What it does not do yet

  • You cannot attach an image to a note.
  • A page is a snapshot. It does not show live data.
  • There is one look: the dark theme.

Documentation

The reference material lives in docs/.

Guide What is in it
Install and quick start Setup skill, Claude Code plugin, Claude Desktop, Cursor, Codex
How it works What the agent writes, the review loop, how it compares with lavish-axi
CLI reference Every command, the rules of a review session, settings
Components and packs What can go on a page, and how to add more
Writing a pack Build and publish your own components
Team review Invite teammates over your own private network
Hosted pages Publish a page and invite reviewers by link
Hub A home page with every review and note; running it all the time
GitHub Action Post a review page on every pull request
Security Who can reach your pages, and how they are protected
Benchmark The speed and cost comparison, with its caveats

Contributing

Bug reports, examples, docs fixes and packs are welcome. CONTRIBUTING.md has the setup and the rules. Issues labelled good first issue are small and have a clear finish line. Security problems: report them privately, not in an issue.

License

MIT. Builds on OpenUI (MIT). The review loop design comes from lavish-axi (MIT).

Yorumlar (0)

Sonuc bulunamadi