Personal AI Agent — FIG. 01

Your own AI agent, on your own machine.

Helm runs on your hardware with full shell, files, web and screen. You message it from Discord, iMessage or your terminal — it does the work and reports back. Owner-locked, private, and powered by any model you choose.

npx github:GOODMAN-PRO/helm
OWNER-LOCKED macOS · WIN · LINUX ANY MODEL $0 SaaS

Sheet 1 / 5 — Thesis

Not a chatbot.
An agent that acts.

Most assistants answer inside a browser tab and stop there. Helm has your terminal, your files, your browser and your screen — and the judgment to use them. Ask for the outcome; it opens the tools, does the work on the machine you own, and brings back the result.

Section A — Capabilities

Everything a real agent needs, on your side of the wire.

Eight things Helm does that a chat window can't. No cloud sandbox, no SaaS account, no data leaving hardware you own.

A1 · core

Runs on any model

Claude on your subscription or an API key — or any other model. Run a free local model with Ollama at $0 and no account, or a hosted one. You pick at install; switch whenever.

A2

Lives on your machine

Full shell, file system and web access — not a locked cloud sandbox. It builds, runs and fixes things for real.

A3

Message it anywhere

DM it on Discord, or iMessage on a Mac, or call it from your shell. No dashboard to babysit.

A4 · always-on

Thinks 24/7, upgrades itself

Background cognition through the day; a nightly self-improvement pass gated by tests that auto-revert on failure. It gets better while you sleep — and never ships a broken change to main.

A5

Owner-locked & private

Only your ID can command it. Your data, keys and conversations never leave hardware you own.

A6

Real long-term memory

A structured memory store plus a self-rewriting Markdown brain that learns your preferences and stays consistent.

A7

Parallel build swarm

Spins up agents in isolated git branches to build and review features at once — smoke-gated, so a broken change never reaches main.

A8

Sees & drives your screen

Screenshots, clicks and types on whichever machine it runs on. It operates apps that have no API when it has to.

A9

Encrypted secrets vault

Hand it credentials safely — encrypted at rest, never echoed into chat, logs or git.

Fig. 02 — Operating sequence

Three moving parts. That's the whole loop.

You don't manage Helm. You talk to it, it acts on your machine, and it answers — one continuous thread across every channel.

step 01 / send

You message

Send a DM from Discord or iMessage, or run helm in a terminal: a task, a question, anything.

step 02 / act

Helm acts on your machine

It runs your chosen model with full tools, right on your hardware — shell, files, web, screen, memory.

step 03 / report

It reports back

You get the result — files built, a problem fixed, a screenshot — straight back in the same chat.

Sheet 2 / 5 — Console

One command: helm.

TERMINAL — macOS · WIN · LINUX zsh · bash · pwsh
# register the `helm` command once
$cd ~/helm; npm link
# then, from anywhere
$helm # open the chat (starts Helm if needed)
$helm "what changed in this repo today?"
$echo "summarize my downloads" | helm

One brain, every channel.

Don't want to open Discord? Talk to Helm straight from your shell. The terminal is a live window into the same running agent — one brain, one memory, one conversation, shared with Discord and iMessage. Type in the terminal and your phone sees it; reply on Discord and the terminal shows it.

Run helm with no arguments for an interactive chat, or pass a message for a one-shot reply. Other commands: helm start, helm setup, helm --help.

Windows tip: in PowerShell, join two commands with ; not && — e.g. cd ~/helm; npm link.

Table 1 — Capability matrix

Built for you, not a data center.

The open-source ecosystem and the hosted players each nail one slice — lightweight, secure, edge, or memory. Helm is the one that does all of it on hardware you own.

How Helm compares to other AI agents across capabilities. Helm column is highlighted.
Open-source & local Managed & hosted
Capability Helm OpenClaw Hermes Agent TrustClaw ZeroClaw PicoClaw NanoClaw Claude Cowork Manus Perplexity Computer Kimi Claw Vellum
Runs on your own machine
Open / self-hosted
Chat from Discord / iMessage~~~
Full shell + file access~~~~~
Clicks & types in any app (no API)~~~~~
Persistent memory + self-rewriting brain~~~~~~~
Background thinking 24/7~
Self-upgrades (auto, tested)
Parallel build swarm~
Encrypted secrets vault~~~~~~~
Owner-locked & private~
Uses your subscription / BYO key~~~~~~
Runs on free / local models~~~
No extra monthly bill~~~~~
Yes ~Partial / limited No

Positioning is based on each project's public description, not an exhaustive benchmark. "Partial" means the capability is limited or hosted-only. Product names belong to their owners. Tell us where we've got a row wrong and we'll fix it.

Fig. 03 — Desktop dashboard

A native window for your Helm.

The desktop client, wired to the same brain as Discord and iMessage — resumable conversations, a live "thinking" status, a background-tasks view, document attachments, and your memory, all in one window.

macOS: Apple Silicon · Intel. Windows: the .exe installer above.

Requires Helm running on your machine (helm in a terminal) — Helm Dashboard connects locally on 127.0.0.1:4625. Builds are unsigned: on macOS, first launch is right‑click → Open; on Windows, More info → Run anyway.

Sheet 4 / 5 — Assembly instructions

One command. Then it's yours.

It brings everything it needs — Node, the Claude Code engine, dependencies — then a wizard walks you through the rest. macOS, Windows and Linux.

BETA
# Works everywhere. Installs everything. You message it from Discord.
$npx github:GOODMAN-PRO/helm

Brings everything it needs: Node, Claude Code (the engine) and dependencies are installed for you. Then pick any model in the wizard; the free/local path (Ollama) needs no Anthropic account at all.

Windows tip: the PowerShell line works even when scripts are "Restricted". If a bare npx says "running scripts is disabled", use the PowerShell line above, or run npx.cmd github:GOODMAN-PRO/helm.

Setup, step by step

The command handles 1–2. The wizard walks you through 3–5.

01

Prerequisites

The installer auto-installs Node 18+ and Claude Code (the engine) if they're missing. Nothing to set up by hand.

02

Fetch & install

Downloads Helm and runs npm install for you.

03

Make a Discord bot

Developer Portal → New Application → Bot → Reset Token. Copy the token.

04

Run the wizard

Paste the token and your Discord user ID, then pick your gateway, your model (including free), and whether to run 24/7.

05

Message it

DM your bot on Discord (or iMessage on a Mac). It acts on your machine and reports back.

Any model, including free

Claude (subscription or API key), or pick a free local model from a list and setup downloads it for you (via Ollama, $0, no account). Hosted models — OpenAI, Gemini, Groq, OpenRouter — work too.

Multiple gateways

A guided setup screen turns on Discord, iMessage, or both. More on the way.

Run it 24/7

Optionally install it as a background service so it's always there.

Create your Discord bot · ~2 min

The one part that's yours to do.

1

New Application

Developer PortalNew Application → name it → Create.

2

Get the token

BotReset TokenCopy. That's your DISCORD_TOKEN. Keep it secret. No special intents needed.

3

Your user ID

Discord Settings → Advanced → Developer Mode on, then right-click your name → Copy User ID. That's OWNER_ID — it locks the bot to you.

4

Add & DM it

OAuth2 → URL Generator → scope bot → open the URL → add to a server. Then DM the bot. Paste both values into the wizard.

Safety & permissions

Helm is a real agent on your machine: shell, files, web, screen, mouse/keyboard, scheduled actions, and a tested nightly self-upgrade. It's owner-locked to your Discord ID, secrets live in an encrypted vault, and it confirms before anything destructive. Start in "ask before each action" mode (the wizard's recommended default) and switch to full autonomy only once you trust it on that machine. Run it on hardware you own. It uses your own Claude account.

Decommission

One line. It leaves your vault and backups alone.

Stops every service, kills the processes, and removes the install (asks first; your HelmBrain vault and backups are left alone).

# macOS / Linux
$curl -fsSL https://raw.githubusercontent.com/GOODMAN-PRO/helm/main/scripts/uninstall.sh | sh
# Windows (PowerShell)
$irm https://raw.githubusercontent.com/GOODMAN-PRO/helm/main/scripts/uninstall.ps1 | iex

Put an agent on your own machine.

No SaaS bill, no account required for the free path, no data leaving your hardware. One command and it's yours.

npx github:GOODMAN-PRO/helm
HELMSHEET 5 / 5DRAWN BY YOUSCALE 1:1