v0.1.1GPL-3.0Windows 10/11 x64

Claude Code, in a real window.

Helm is a Windows desktop app for people who already run Claude Code, usually several sessions at once. It keeps the CLI you installed and replaces the terminal screen with a sidebar of every session, a conversation you can read, and panels for agents, changes and usage.

Download for Windows View on GitHub

Helm-Setup.exe · 109.2 MB · Windows 10/11 x64 · v0.1.1

Needs your own Claude Code install, signed in.

Fix the login redirect loop

Running npm test · 3m 04s 2 agents 21% context storefront

Signing in sends people straight back to /login. Find out why and fix it.

The guard reads the session before the cookie is written, so a fresh sign-in looks signed out and gets sent back.

Readsrc/auth/guard.js48 lines

Editsrc/auth/guard.js+4−1

−if (!session.user) return toLogin(req);

+await session.ready;

+if (!session.user) {

+ return toLogin(req, req.path);

Bashnpm test -- auth12s

Helm’s window, rebuilt from its own interface code. Sample data, not a real session.

How it works

It’s the real Claude Code.

Helm doesn’t reimplement Claude. Every session is the Claude Code CLI you install yourself, driven through Anthropic’s Claude Agent SDK. Helm only replaces the terminal screen.

Your setup carries over.: Your CLAUDE.md files, memory, hooks, MCP servers and skills work exactly as they do in the terminal.
Your sign-in stays yours.: Helm never signs you in and never sees your login. Sessions use whatever Claude Code is signed in with, or an Anthropic API key you add in Settings.

The screenHelmSidebar, conversation, inspector, phone page
Yours, installed by youClaude Code CLICLAUDE.mdmemoryhooksMCP serversskills
YoursYour sign-inYour Claude account, or an Anthropic API key

Sessions

Run several sessions, and read every one.

Every session at a glance.
Every project and session in the sidebar, with a dot for working, waiting on you or error, unread counts, filter chips, and a split view of two sessions.
Read what Claude is doing.
Markdown answers, collapsible thinking, tool cards with diffs, and permission, question and plan cards you answer in place.
Find your way around.
Search and resume past sessions, a Ctrl+K palette, a quick-prompt hotkey (Ctrl+Alt+Space), and tray notifications.

Inspector

What the terminal scrolls past, kept in view.

The inspector sits beside every session: the sub-agent tree with context bars, Changes with diffs and checkpoint rewind, Files, a localhost Preview, and Usage for your 5-hour window, week and cost.

The 5-hour and week readings exist only on Claude account plans. With an API key, Usage shows cost.

peak 2 at once · $0.88 · 1 past 32% · session 21%

Trace where the session cookie is written2m 40s

Explore · Opus 5.514 tools · Grep$0.41

27.4%
Run the auth tests against the fix1m 12s

general-purpose · Opus 5.59 tools · Bash$0.28

34.1%

Past 32%, due for a hand-off.
Find every redirect to /loginran 1m

Explore · Opus 5.5ended 4m ago6 tools · Read$0.19

12.0%

Helm’s inspector at its default 360 px, rebuilt with sample data.

Overnight queue · Production guard

Queue the work for tonight. Pause before production.

Overnight queue.

Add prompts to a queue. Each one starts when the 5-hour budget allows, or after a time you set. It won’t start while your 5-hour usage is above the share you set.

Sample jobs; the cap, start times and job lengths are examples, not measurements.

Production guard.

Helm can stop Claude before it pushes to a production branch you list or runs a Railway deploy, and ask you first. It holds in every permission mode. It’s opt-in: you list the repos in prod-guard.json.

Production push · My App mainPushing My App main deploys production.Production

Helm’s production guard stopped this, whatever the permission mode. Nothing runs until you choose; no answer in 30 min counts as Deny.

Commit the fix and push it

git push origin main

%APPDATA%\Helm\prod-guard.json

{ "repos": [ { "name": "My App",
    "remotes": ["github.com/your-org/your-app"],
    "branches": ["main"] } ] }

The example config from Helm’s README, and the card it produces, rebuilt with a sample command.

Phone

Answer it from your phone.

The same UI in a phone layout, served from your PC over Tailscale. You pair once by scanning a QR code.

It’s a web page on your phone, not an app you install.

Phone

Scan with your phone’s camera, or open the link on any device on your tailnet. The link carries Helm’s key: treat it like a password.

https://your-pc.your-tailnet.ts.net/?key=…Copy

Fallback if https is down: http://your-pc:4871/?key=… Copy

Allow phone access

Helm’s pairing card in Settings, rebuilt. Sample links; the QR code is a drawing that encodes nothing.

HelmNew

Filter sessions

Working2Waiting2Error1

storefront

Fix the login redirect loopRunning npm test · ~14mnow

Release notes for v1.3.2Stalled · 12m12m

booking-portal

Customer import previewWaiting for your answer2m

mobile-app

Bundle size report on releaseWaiting for permission1m

api-server

Scheduler clock driftAPI error: overloaded1h

Home2SessionsHistoryQueueSettings

Helm’s phone layout, which it switches to below 821 px wide, rebuilt with sample data.

Download

Download and set up.

Helm

Claude Code, in a real window.

Windows 10/11 x64 · v0.1.1 · Helm-Setup.exe, 109.2 MB

Needs your own Claude Code install, signed in to your Claude account, or an Anthropic API key.

SHA-256 SHA-256 c1f0c7b17e859f1e982652cf959a4f83e156fcac0d2547d3b644e958ddb8536ac1f0c7b17e859f1e…d3b644e958ddb8536a

Download for Windows Release notes

After you download

Install Claude Code from Anthropic (code.claude.com/docs), open a terminal, run claude and sign in once.
Run Helm-Setup.exe.
Windows shows “Windows protected your PC” because the installer isn’t signed yet. Click More info, then Run anyway. There’s no admin prompt.
A short install window appears, then Helm opens.
First run only: a “Getting Helm ready” window downloads Anthropic’s Claude Agent SDK (about 1.6 MB) from the npm registry. If that fails, it says why (offline, npm blocked, a proxy), and Try again retries.

Updates. Settings → Helm → Updates shows when a new version is out. Choose Download the installer and run it over the old one; your settings stay.

Uninstall. Settings → Apps → Helm → Uninstall. It removes the app and the downloaded SDK. Your settings and sessions in %APPDATA%\Helm stay.

Requirements

Windows 10 or 11, x64. Windows only.
Your own Claude Code install, signed in to your Claude account, or an Anthropic API key.
Internet on first run, to fetch the SDK.

Why the SDK isn’t in the installer

The Claude Agent SDK is Anthropic’s software, so Helm’s installer doesn’t ship it. Helm fetches it from npm on your machine the first time it runs, and the build fails if any Anthropic file ends up inside.

Prefer to build from source?

You’ll need Node.js 22 or newer, with npm, and git.

$ git clone https://github.com/atlas-studio-software/helm
$ cd helm
$ npm install
$ npm start

node tools/make-exe.js builds a Helm.exe with a proper icon, and two scripts in bin\ add a Start Menu entry and a helm command.

Phone, optional: Tailscale on the PC and the phone, plus a one-time firewall script run as admin.

Limits

What Helm doesn’t do yet.

It runs on Windows only. There’s no macOS or Linux build.
It doesn’t include Claude Code or sign you in. You install Claude Code yourself.
The phone side is a web page over Tailscale, not a phone app.
The installer isn’t code-signed yet, so Windows warns you before it runs.
It has been built and tested on one machine.

Source and licence

Version 0.1.1, GPL-3.0. Issues are welcome; pull requests aren’t accepted right now.

Helm is an independent project. It is not affiliated with or endorsed by Anthropic. Claude and Claude Code are trademarks of Anthropic.