Open this on your laptop

go.fzj.de/2026-hdc-rs

Resources:

• This document
• Source code of this talk

Team:

Who Is This For?

• You’re a researcher who uses AI tools (ChatGPT, Claude, etc.) but want more power
• You’re tired of copy-pasting between tools and want automation
• You need privacy — your data can’t leave your machine or Helmholtz infrastructure

You don’t need:

• Prior experience with AI agents
• Deep technical knowledge (terminal familiarity helps)
• A powerful machine (local models optional — Blablador API works on any laptop)

Good to know:

• WHAT YOU ARE DOING
• You’re giving a powerful AI agent access to your tools and data
• It can read/write files, execute code, browse the web, and more
• With great power comes great responsibility — be mindful of what you ask it to do

AI Agent?

• A program that uses an LLM to plan and execute tasks using tools
• Benefits for researchers: automation, reproducibility, speed

Why Hermes Specifically?

Hermes vs. ChatGPT in a browser

Hermes

• Remembers across sessions
• Writes its own reusable skills
• Prunes them in the background
• And validates them offline through an evolutionary engine called GEPA
• Open-source, multi-model, tool-calling, privacy-first
• Runs locally on your machine — your data stays with you
• Connects to Blablador
  • Guide: sdlaml.pages.jsc.fz-juelich.de/ai/guides/blablador_hermes
• Also supports: OpenRouter, Claude, ChatGPT, Gemini and local models

Hermes vs. other agent frameworks:

• Simpler setup — curl | bash vs. Docker + config hell
• Built-in skills — 50+ reusable workflows out of the box
• Privacy-first — runs locally by default, Blablador integration
• Research-focused — arxiv, llm-wiki, citation tools bundled

Why Blablador?

“So hoping that in 1 year one can use JUPITER or JARVIS to run something on the level of Claude of now is entirely unrealistic.    Like, fairy tale unrealistic.”

• From a colleague last week.
• I hope he is wrong.

Life can be hard even if you are Microsoft

Remember

• When you stop paying the provider, the tokens stop
• Your work stops.

Why Blablador?

• Blablador API: state-of-the-art open models, 35 GPUs 24/7, Helmholtz authentication
• More coming
• Explosion in token numbers
  • Dell CEO talks about 320x increase since 2025
  • Commercial cloud provider strategy unsustainable for whole Helmholtz

It’s all about the right tool for the job

Why NOT Blablador?

• User asked me yesterday (11.06.2026):
  • Why is your infrastructure not as good as Claude or ChatGPT?
  • Chatgpt: 600 billion dollars in infra plus people.
  • Claude: 30 billion dollars in infra plus secret accords.
  • Blablador: 1 person, 50 GPUs, 0 dollars in infra costs (all donated)

🤌

🖕

Current OUTDATED

(11.06.2026) state-of-the-art models

Closed vs Open Source, result of 10 benchmarks

Prerequisites

• Terminal access (macOS, Linux, or WSL on Windows)
• Telegram (if desired)

Recommended but not required:

• Basic Python familiarity
• Git installed
• Telegram account (for bot integration)

Outline

• Installation & setup: 20 min
• Configuration & model selection: 10 min
• Telegram bot: 10 min
• Skills (arxiv, llm-wiki): 30 min
• Custom skill creation: 10 min

Configuration

Blablador Token:

• Log in to codebase.helmholtz.cloud
• Edit Profile → Access → Personal Access Tokens → Add New Token
• Check date, add scope “Read user”
• Copy the token - you won’t see it again.

Installing Hermes

curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash

• hermes setup — Full setup
• Custom endpoint (enter URL manually)

Connecting to Blablador

• API endpoint: https://api.helmholtz-blablador.fz-juelich.de/v1/
• Still have your token?
• Api mode: 1 - Auto-detect
• Default model: alias-fast (Gpt-OSS-120b)

Connecting to Blablador

• Terminal Backend: local
• Most other options depend on what you want to do — We can change them later

Connecting to Blablador

• Verify: hermes doctor

Memory

🍆

Size matters

• (Context size, that is)
• Bigger context windows → more memory, less forgetting
• One should balance model size, context window, and latency for the task at hand

Creating the searxng-search Skill

Prompt to create it:

"Add a custom search backend to Hermes. I have a self-hosted SearXNG instance at https://search.blablador.fz-juelich.de. Remember to use it for ANY web search."

Quick Test

Verify everything works:

hermes chat -q "What is the capital of France?"

Blablador Model Aliases

Blablador Model Aliases

Blablador Model Aliases

• Aliases are stable — the actual model can change without breaking your code
• All models free for Helmholtz researchers
• Important: Context windows are finite (8K-256K). Rate limits apply. Monitor usage via hermes logs --follow

Practical: making it more than a chatbot

Creating a Telegram Bot

Register a bot: message BotFather on Telegram, then add to ~/.hermes/.env:

TELEGRAM_BOT_TOKEN=YOUR_TOKEN
TELEGRAM_CHAT_ID=<your_chat_id>

Run: hermes gateway run and send /status in your Telegram chat.

Bots

Why not whatsapp or matrix?

• Telegram: Bot API is well-documented and supported by Hermes out of the box.
• Matrix: integration is possible but requires more setup and you need multiple accounts or platform bot support (fz doesn’t).
• Whatsapp: no bot API — so it’s not a good idea to run on your own account

Voice & TTS

hermes tools enable tts
hermes config set tts.provider NeuTTS

• Generate speech: /voice or /speak "Hello class"
• Fallback: hermes config set tts.provider edge

Core Hermes Commands

Core Hermes Commands

Hermes Profiles

• Each profile has its own model/provider, config, memory, and tool approvals
• hermes profile list
• A profile becomes a new tool:
  • hermes profile create researcher --description "Reads source code and external docs, writes findings."
• researcher help

Goals

• /goal: a standing objective that survives across turns.
• After every turn a lightweight judge model checks whether the goal is satisfied by the assistant. Default: 20 turns
• Ralph Loop
• Example: /goal Create four files /tmp/note_{1..4}.txt, one per turn, each containing its number as text

Samba picks the right skill

Key Skills for Research

Skills are reusable, documented workflows — load with /skill <name>, then ask natural language questions.

• hermes-agent — Full setup guide, CLI reference, troubleshooting
• arxiv — Search papers, fetch abstracts, generate BibTeX entries
• llm-wiki — Build a personal, inter-linked markdown knowledge base
• plan / test-driven-development — Development workflows

Skills

Load a skill before related work:

/skill hermes-agent
/skill arxiv
/skill llm-wiki

• Full list: hermes skills list or hermes skills browse

Custom Skills: Creating Them

hermes-tech-digest — daily tech news digest to Telegram:

"Add a skill for generating a daily tech news digest and sending it
to Telegram via the gateway. Remember to use it whenever someone wants a daily news roundup."

Write your own: /skill hermes-agent-skill-authoring

hermes-tech-digest: Configure & Run

Tell Hermes your interests:

"Customize the hermes-tech-digest skill. Set my topics to:
AI, machine learning, open source, Apple Silicon"

Run once to test: load the skill, then ask Hermes to run it.

Schedule daily delivery:

"Scheduled hermes-tech-digest to run every day at 8 AM"

arxiv Skill: Literature Search

/skill arxiv

Now ask natural language questions: - “Search for papers on reinforcement learning for scientific discovery” - “Find the top 5 papers on transformer architectures and generate BibTeX” - “Save the BibTeX entries to ~/papers/rl_discovery.bib”

Key point: Skills change the context — you ask natural language, not special commands.

Practical Workflows

• Literature search with arxiv skill
• Basic scripting and data analysis
• Telegram bot setup
• Code & Development: Scripting, debugging, deploying services
• Experiment Management: Cron jobs, background tasks, parameter sweeps

Practical: Data Analysis & Scripting

"write a script that fits a Gaussian to column 3 of my_data.csv"

• Always use uv for package management — never pip:

  uv pip install pandas matplotlib scipy

• Debug with Hermes: “Explain the error in my script”

• Reproducible: “Create a requirements.txt for this project”

Practical: Experiment Management

hermes cron create "0 2 * * *" "run_training.py --nightly"

Monitor running jobs:

hermes logs --follow

Stop a job:

/stop

or

hermes cron pause <job_id>

Important: /background runs in the current session. hermes cron creates persistent scheduled jobs.

Practical: Writing & Publishing

Convert notes to LaTeX:

"convert my notes in research_notes.md to LaTeX"

Check citations: “Verify all BibTeX entries are correct”

Send to Telegram: “Send the PDF to my class Telegram channel”

Practical: Voice Announcements

• Generate lecture audio with VibeVoice:

  /voice on
  "explain attention mechanism" --voice

• Great for accessibility, language support, and review

• Example: “Create a 2-minute summary of today’s lecture”

Common Issues & Troubleshooting

• Model won’t load:
• Blablador: Check API key in ~/.hermes/config.yaml
• Local MLX: Ensure enough RAM (8GB+ for 7B models)
• Try different alias: /model alias-large --provider blablador

Common Issues & Troubleshooting

• Tool permissions denied:
• Check Hermes config for approvals
• Set to smart for auto-approval of low-risk actions
• Review ~/.hermes/config.yaml for tool restrictions

Common Issues & Troubleshooting

• SearXNG skill not working:
• Test endpoint: curl https://search.blablador.fz-juelich.de/api?q=test
• Re-create skill with correct URL

Common Issues & Troubleshooting

• Telegram bot offline:
• Check token: hermes config show
• Restart gateway: hermes gateway run
• Verify chat ID format (no leading zeros)

Common Issues & Troubleshooting

• Tools not working after enabling:
• Critical: Tools require /reset to take effect
• Run /reset after hermes tools enable <name>
• Tool changes do NOT apply mid-conversation (preserves prompt caching)

Common Issues & Troubleshooting

• Skill loaded but nothing happens:
• You must ask natural language questions after loading
• Example: /skill arxiv then “search for papers on X”
• Skills change context, they don’t add commands

Security & Best Practices

• https://github.com/multica-ai/andrej-karpathy-skills/blob/main/CLAUDE.md

Security Configuration

hermes config set security.redact_secrets true
hermes config set approvals.mode smart
chmod 700 ~/.hermes && chmod 600 ~/.hermes/.env

• Never share your .env or API keys
• Review commands before approving destructive operations
• Use smart mode - auto-approves low-risk, prompts on high-risk

Privacy & Data Protection

• Built-in memory stores data locally in ~/.hermes/memories/
• External providers (optional) - all self-hosted, open-source
• Secret redaction - automatically hides API keys in tool output
• No data exfiltration - private data stays private

Privacy & Data Protection

Verify your setup:

hermes config check
hermes memory status

What I hope you can do with Hermes

• Automate literature reviews — search arxiv, extract summaries, build citations
• Build personal knowledge bases — link concepts, track research evolution
• Write & debug code faster — let agents handle boilerplate, you focus on logic
• Schedule recurring tasks — daily digests, nightly training, automated reports
• Communicate via Telegram — query your agent on-the-go, get voice responses
• Create custom skills — encode your own workflows for reuse

Next Steps

• Deepen your skills:
• Explore hermes skills browse for 50+ native skills
• Read /skill hermes-agent for full CLI reference
• Join the Hermes Discord: https://discord.gg/NousResearch

Next Steps

Advanced topics: - Multi-agent orchestration with /delegate_task - Custom tool development (Python SDK) - Deploying Hermes as a service for your team

Samba and Potato have questions too

Questions?

• Slides: https://strube1.pages.jsc.fz-juelich.de/hermes-course
• Source: https://gitlab.jsc.fz-juelich.de/strube1/hermes-course
• Hermes docs: https://hermes-agent.nousresearch.com/docs/