Zero to autonomous in about thirty minutes.

• A credit card for the VPS subscription. ~$5/mo at the cheapest tier; annual is fine.
• A Telegram account on your phone or desktop. Free.
• An LLM API key. Anthropic, OpenRouter, or local Ollama. Pick one to start.
• An image/video API key if you want creative output. KIE.ai, Higgsfield, or both.
• 30 minutes of focus. The install runs unattended, but you'll context-switch a few times.

Hermes is one program running on a small server you rent. You talk to it through Telegram (or Discord). It reads and writes from a small local database (its memory), calls an LLM for thinking, and uses tools to do things in the world — generate images, push code, send messages, run code, post to X.

The whole thing is one Docker container running on the VPS. You don't need to write code to install it. The install path uses Hostinger's one-click VPS template.

Go to Hostinger and buy a KVM 2 plan or higher. KVM 1 will technically work for a single user but the agent feels sluggish at scale. KVM 2 (~$8/mo or $5/mo annual) is the sweet spot.

Pick a region close to you (latency matters for Telegram round trips). Skip every Hostinger upsell except basic backups.

Inside Hostinger's hPanel:

1. Open your VPS dashboard.
2. Hit Operating System → Templates.
3. Search Hermes in the template gallery. Click Select.
4. Confirm and let it deploy. Three to five minutes.

When the template lands, you'll see a green "Hermes Gateway running" status in your VPS panel. The Gateway is the front door — a tiny HTTP server that listens for messages from Telegram and routes them into the agent.

# Optional — most people skip this step
ssh root@your-vps-ip
systemctl status hermes-gateway
# Should print: Active: active (running)

In Telegram, search for the user @BotFather. Open the chat and send /newbot.

BotFather will ask for two things:

• A display name — the name you'll see in chats. e.g., "Hermes for Kyle"
• A username — must end in _bot. e.g., kyle_hermes_bot

BotFather replies with a token that looks like this:

7842931572:AAH…BqXcVnLm9PoQrStUvWxYz

Paste the token into Hermes' settings panel under Telegram → Bot Token. Hit save. The Gateway will pick it up within ~10 seconds and start polling.

In Telegram, search @userinfobot, open the chat, and send /start. It replies with your numeric user ID — a 9-to-10-digit number like 1759384210.

Add that number to the Hermes Allowed Users list. You can add multiple — one per teammate later.

You have three sane choices:

For 95% of people: Anthropic. Set claude-opus-4-7 as the default model and load $20 of credits. Auto top-up when balance falls below $5.

PROVIDER=anthropic
DEFAULT_MODEL=claude-opus-4-7
FALLBACK_MODEL=claude-haiku-4-5-20251001
ANTHROPIC_API_KEY=sk-ant-…your key…

The LLM gives Hermes a brain. These keys give it hands for visual work.

• KIE.ai — fast, cheap image generation. GPT-Image-2 and Nano-banana-2 are both available. Best default for thumbnails and carousels.
• Higgsfield — short-form video generation, avatar lip-sync. Pricier per generation but high-quality.
• ElevenLabs — voice cloning, TTS, dubbing. Wire it if you want voice replies or auto-narration.

KIE_API_KEY=kie-…
HIGGSFIELD_API_KEY=hf-…
ELEVENLABS_API_KEY=el-…
DEFAULT_IMAGE_MODEL=gpt-image-2

Open Telegram. Find your bot (the username from step 3). Send:

Hi. What can you do?

Hermes will think for a moment, then reply with a self-inventory — a short list of capabilities based on what you've connected so far. If you only have an LLM key, it'll list memory, code, web research, and creative writing. As you add more keys (steps 6, 10, 11), the list grows.

If nothing happens after 30 seconds, jump to troubleshooting. The two most common causes are a misspelled bot token and a missing user-ID allowlist entry.

Hermes can route prompts to different models based on context. You set a default, plus rules.

DEFAULT_MODEL=claude-opus-4-7
CODE_MODEL=claude-sonnet-4-6
CLASSIFY_MODEL=claude-haiku-4-5-20251001
ROUTER=on            # auto-pick model per task

Hermes loads soul.md at the top of every conversation. Anything in here is permanent context — projects, voice, do's and don'ts, people in your life, your business model.

# Soul · Kyle Painter

## Who I am
- CEO at Disruptors Media (AI implementation agency, US + PH team)
- I think in systems and content velocity
- I prefer demos over explanations, defaults over options

## How I write
- Plain language, no jargon, no emojis, no em dashes
- Short sentences. Active voice
- If you suggest something, suggest one thing, not three

## My current projects
- Hermes onboarding for the Disruptors team
- Daily content (M/W/F at 6 PM MT)
- SEO-GEO Tool v2 with Bryan

## My team
- Bryan (engineering lead, PH)
- Rio, Camilo, Kimball (account managers)

## Rules
- Never DM clients without my approval
- Always confirm before deploying or sending money
- If a request looks unusual, ask first

Open Hermes' Integrations panel. You'll see OAuth buttons for Gmail, Google Calendar, Google Drive, Notion, and Dropbox. Each is a one-click connect.

Suggested order:

1. Calendar (read-only) first — lets Hermes see your day. Low risk, high signal.
2. Drive / Notion (read-only) — gives it your knowledge base.
3. Gmail (draft-only) — drafts replies for you to send. Magic for inbox triage.
4. Calendar (read/write) after a week of read-only.
5. Gmail (send) last, behind the approval flow we set up in step 15.

This is where Hermes goes from "answers your questions" to "ships things." Add as many or as few as you want.

• X / Twitter — bring an X API key. Hermes drafts and posts on your behalf.
• WhatsApp Business — via the Meta Cloud API. Hermes can send templated messages, never reads inbound unless you opt in.
• Vista Social or Blotato — multi-platform schedulers. One write, ten platforms.
• Philips Hue / Home Assistant — smart home routines. Yes, you can ask Hermes to dim the lights.
• GitHub — fine-grained access token scoped to specific repos. Hermes can open PRs, write issues, run workflows.
• Stripe / PayPal (read-only) — revenue dashboards on demand. Never give it write access; nothing should auto-charge.

Try this prompt verbatim in Telegram:

Spawn three sub-agents:
1. Pull the top 5 AI headlines from the last 24h.
2. Draft a carousel for each in my voice (use my Soul).
3. Generate a thumbnail for each with KIE.

Report back when all three are done.

Hermes will fan out into three parallel workers, each with its own scratchpad memory. When all three return, the parent agent stitches the result back to you.

You'll see status updates ("Agent 2: drafting carousel 4 of 5…") in the same thread. The whole thing finishes in 4–6 minutes for a five-headline batch.

Add up to 3 device-bound users by default (this is the IP-based access cap; raise it with a setting flag if you trust your VPN posture).

USERS=
  - id: 1759384210     # Kyle
    soul: soul-kyle.md
  - id: 2103948271     # Bryan
    soul: soul-bryan.md
  - id: 3140293847     # Rio
    soul: soul-rio.md
MAX_DEVICES_PER_USER=3

Each user has isolated memory. Bryan's projects don't leak into Rio's context. Same agent, different brains.

Open the Hermes dashboard at https://your-vps-ip:8080 (do NOT bind to 0.0.0.0 — keep it on localhost or behind Cloudflare Tunnel). Cost panel shows tokens per model, per user, per day.

DAILY_USD_SOFT_CAP=5     # warn at $5/day
DAILY_USD_HARD_CAP=15    # pause new requests at $15/day
MONTHLY_USD_HARD_CAP=200  # circuit breaker for the month
PER_USER_MONTHLY_CAP=50   # in multi-user mode

Some tools — sending email, posting to X, executing arbitrary code, deleting files — can't be undone. Tag them as requires-approval and Hermes will pause and DM you a confirmation card before firing.

REQUIRES_APPROVAL=
  - gmail.send
  - x.post
  - whatsapp.send
  - github.push_to_main
  - shell.execute
  - file.delete

The approval card is a tap-to-confirm message. You see exactly what the agent's about to do, with the input arguments rendered. Tap Approve, the action runs. Tap Reject, the agent retries with feedback or stops.

Three things to set and forget:

1. Backups. Hermes' memory lives in /var/hermes/data/. Mount a Hostinger volume snapshot, schedule daily, retain 7. ~$1/month.
2. Updates. Hermes self-updates by default (NOS Research ships often — Kanban, level 7, etc.). If you want to pin a version, set AUTO_UPDATE=off and bump manually.
3. Uptime. Add a UptimeRobot or Better Uptime check on /health. Free, pings every 5 min, alerts to your phone if the gateway dies.

AUTO_UPDATE=on
BACKUP_PATH=/mnt/snapshots/hermes
BACKUP_RETENTION_DAYS=7
HEALTH_ENDPOINT=/health
LOG_LEVEL=info

• Rotate every key once after install. Whatever you pasted during setup, generate a fresh one and replace it. Removes paste-history exposure.
• Bot token is not in any chat or file you'd ever share. Live in a password manager only.
• Allowlist is locked. Only your Telegram user IDs. No "anyone in the group."
• Workspace integrations start read-only. Upgrade scopes per integration only after a week of use.
• Approval flow is on for any tool that sends, posts, or deletes. Step 15.
• No payment processor has write scopes. Read-only or not connected.
• Spend caps are set. Soft + hard, daily + monthly. Step 14.
• Dashboard is bound to localhost or Cloudflare Tunnel. Never 0.0.0.0.
• VPS root login is key-based, not password-based. Disable password SSH in /etc/ssh/sshd_config.
• Backups are running and tested. Restore one before you ever need to.

1. Run the "earn back the VPS" prompts. Six of them are on the lead-magnet page. Pick one. Watch it execute. That's the dopamine hit that locks the habit in.
2. Rewrite your Soul. First draft is always too generic. After a day of use, you'll know the specific things you wish you'd told it. Add them.
3. Tell one friend. The fastest way to lock learning is to teach. Send them this guide and walk them through the install in 30 minutes on a call.

Want help? Comment HERMES on the post you came from, or reach out at kyle@disruptorsmedia.com.