Run a whole fleet of Claude Code agents without babysitting them. HELM watches every session, approves the routine "Allow" prompts that freeze them, and sends only the risky calls (delete, send, publish, pay) to your phone. Nervous about a tool that clicks for you? Flip on Observe mode and watch it for an hour before it ever touches anything.
This is HELM running. The routine prompts get approved in seconds. The irreversible always waits for you.
Routine prompts: approved in ~4s, even minimized. Delete · send · publish · pay → always you.
You kick off five Claude sessions, walk away, and come back to find every one of them frozen on "Allow Claude to edit this file?" — waiting on a click you could have made in your sleep. The promise of autonomous agents dies on a permission dialog. HELM is the hand on the wheel that keeps them moving.
HELM clicks the real buttons in Claude Desktop through Windows accessibility — no screen-scraping, no fake keystrokes, no runaway processes.
Permission prompts like "Allow Claude to edit X" get approved in about four seconds — for the session on screen and the ones running in the background.
Anything irreversible goes to Telegram with one-tap options. You decide; HELM applies it and remembers it for next time.
A session that finishes and sits idle gets nudged to continue. Opt-in per session, rate-limited, and never your private chats.
It scores its own accuracy against your choices and proposes new rules you accept with a tap. It needs you a little less every week.
A heartbeat + watchdog recover it from hangs, not just crashes. It self-tests every 20 minutes and pings you the moment anything degrades.
Text it /status, /fleet, /pause, /resume. See every machine and what's waiting, from anywhere. Shrink it to a tiny strip when you don't need the panel.
An early version once spawned background processes and took a machine down. That mistake became a permanent, self-verified rule. Here is exactly what HELM will and won't do.
Auto-approves routine tool permissions — edit a file, run a read-only command, continue a task.
Escalates to you any question it isn't confident about, with your options one tap away.
Never auto-approves the irreversible — deleting, sending email, publishing live, posting to social, or anything involving payment. Those always wait for you.
Never spawns background AI processes. The pattern that caused the original failure is disabled and continuously verified by a built-in self-test.
Acts only on what's real. Every action is a no-op unless the actual on-screen button exists — it can't fire into the wrong window.
Stays on your machine. No cloud account, no data leaves your PC. Telegram control is optional and uses your own bot.
Reads every live Claude session and what each is waiting on.
Routine → approve. Ambiguous → choose smartly. Risky → you.
Clicks the real button. Verifies the session actually moved.
Logs it, learns from it, shows it on your phone.
No subscription, no account, no telemetry. The free tier does the thing everyone wants. Pro unlocks the unattended-operator features.
HELM is built specifically for Claude Desktop on Windows (including Claude Code sessions run inside it). It drives Claude's real interface, so it does not work with ChatGPT, Gemini, or other assistants today. If you live in Claude all day, it's built for exactly you.
No. The Windows download is a one-click HELM.exe — no Python, no setup wizard. Unzip it and double-click. You only need Windows 10/11 and Claude Desktop. (A source build is also available for developers who want to tinker.)
Free auto-approves the routine prompts that stall your sessions — the thing you came for. Pro ($39 once) adds the unattended-operator features: phone control & escalation, stall revival, learning that proposes its own rules, multi-machine fleet view, and smart decisions. Upgrade right inside the app.
No. HELM never auto-approves irreversible actions — deleting, sending email, publishing, posting, or anything with payment always waits for your explicit approval. It only ever clicks a button that's actually on screen, and a built-in self-test verifies its safety rules every 20 minutes.
No. HELM runs entirely on your PC. There's no cloud account and no telemetry. The optional phone control uses a Telegram bot you create and own.
No, it's included. You connect your own free Telegram bot; HELM talks to it directly. Optional smart decision-making can use a cheap third-party model (OpenRouter, DeepSeek, or OpenAI) with your own API key, and it works fine without one.
Email me within 14 days and I'll make it right or refund you. It's a tool meant to save you hours; if it doesn't, you shouldn't pay for it.
Download it, double-click HELM.exe, and close your laptop knowing the work keeps moving. No Python, no setup wizard — it just runs.
Windows 10/11 · Claude Desktop required · one-click .exe · ~37 MB · Free tier — upgrade to Pro in-app