Most AI agents forget everything the moment you close the tab.
Next session, back to zero. You explain context again. It starts over. Every time.
Hermes works differently. It saves what it learns. Every task you give it, it writes the procedure to a file on your disk. Next time you run something similar, it finds that file and uses it. A month in, your Hermes has 30-50 of these skills sitting in a folder. It gets faster. It gets more accurate. It compounds.
I set it up on a regular laptop. No special hardware. Took about 30 minutes. First week I closed three clients at $300-400 each doing competitor research reports. Actual work per report: 15 minutes.
Here's the full setup.
What Hermes is
Open-source agent framework from Nous Research. 140,000 GitHub stars in three months. Most-used agent on OpenRouter right now. NVIDIA featured it in a May blog post running on their new DGX Spark workstation.
You don't need that hardware. A MacBook with 16GB RAM works. So does any Windows machine with a mid-range GPU.
Three folders on your disk do all the work:
1~/.hermes/memory/ your preferences, projects, patterns2~/.hermes/sessions/ indexed history of everything3~/.hermes/skills/ learned workflows saved as .md files
That skills folder is the whole point. Agents with 20+ self-created skills complete similar tasks 40% faster than a fresh instance. Not better output. Less time to get the same result.
The service
Competitor research reports for early-stage startups and small SaaS companies.
A founder wants to know what their three main competitors are doing. Pricing, positioning, what customers hate about them, where the gaps are. Normally that's 3-4 hours of work for someone. I charged $300 and delivered same day.
Hermes does the actual research in 15 minutes.
What most people pay:
1Service Cost2─────────────────────────────────────3Freelance analyst $150-3004Research firm (minimum) $500-20005DIY 3-4 hours of your time
What this costs:
1Tool Cost2─────────────────────────────────────3Hermes Agent $04Ollama $05Qwen 3.6 27B model $06Your laptop $07Electricity ~$2/month8─────────────────────────────────────9Total $0-2/month
Setup (30 minutes)
Step 1. Local model server
Go to lmstudio.ai. Download and install it.
Open LM Studio, go to the Discover tab, search Qwen 3.6 27B. Pick Q4 quantization. Download takes 10-15 minutes.
After that: Developer tab, load the model, enable "Serve on Network" in settings, hit Start Server. Runs on:
1http://localhost:1234
Open that URL in your browser. If you see JSON, it's working.
If you prefer terminal, use Ollama:
1ollama pull qwen3.62export OLLAMA_HOST=0.0.0.03ollama run qwen3.6 -c 65536
That -c 65536 flag is not optional. Ollama defaults to 4K context. Hermes needs 64K. Skip it and nothing runs.
Step 2. Install Hermes
1bash scripts/install.sh23source ~/.bashrc45hermes --version
Get the install script from: github.com/NousResearch/hermes-agent
Windows users run this inside WSL2.
Step 3. Connect to your model
1hermes model
Pick "Custom endpoint" from the menu.
1URL: http://localhost:1234/v1 (LM Studio)2 http://localhost:11434/v1 (Ollama)3API Key: leave blank, press Enter4Model name: exact filename from LM Studio, or "qwen3.6" for Ollama
If you get "Model context too small" at startup, go back to your model server and set context to 65536. This is the most common problem. Fix is always on the model server side.
Step 4. First session
1hermes
Paste this as your first task:
1Research three competitors for a project management tool targeting2freelancers. For each: positioning, pricing, top customer complaints3from reviews, one gap in their offering. Save this as a skill so we4can reuse the process next time.
Hermes breaks it into subtasks, searches, writes the report, saves the procedure to ~/.hermes/skills/. Next research task runs faster because the skill is already there.
Type /exit when done.
Step 5. Check it worked
1ls ~/.hermes/skills/
You should see .md files. Open one. It's a structured workflow with steps and notes. That's Hermes learning.
Empty folder means the install didn't finish. Re-run the script.
Telegram gateway
1hermes gateway
Pick Telegram. Go to @BotFather, create a new bot, paste the token.
Now you can text your agent from your phone while the laptop runs at home. Changes how it feels completely.
Finding clients
Three places that worked week one:
Upwork. Search "competitor analysis" or "market research." Filter by last 7 days. Send 10-15 short messages per day. Offer to send a sample report. Build the sample with Hermes before you have any clients.
X/Twitter. Search "anyone know" + "competitor research." Founders post this constantly. Reply, offer a sample, don't pitch.
Cold email. Go to Product Hunt, filter launches from last 30 days. Email the founder directly. One sentence, link to sample. Subject: "quick competitor research for [product name]."
First client usually comes in 3-5 days if you're sending enough messages.
The math
1Week 12─────────────────────────────────────3Setup 2 hours4Outreach per day 1 hour5Reports delivered 36Revenue $900-1,2007Work per report 15-20 min
1Month 12─────────────────────────────────────3Reports sold 10-154Revenue $3,000-4,5005Retainers started 2-36Monthly recurring added $600-900
1Month 32─────────────────────────────────────3Skills in ~/.hermes/skills/ 30+4Time per report 10 min5Retainer clients 6-86Monthly recurring $1,800-2,4007One-off reports $1,500-2,0008Total $3,300-4,400/month
Common problems
"Model context too small" at startup. Set context to 65536 on your model server. This is 80% of all setup issues.
Hermes is slow. Drop from 35B to 27B model, or Q6 to Q4 quantization. CPU-only means 2-3 minutes per response. Get a GPU or use the cloud API.
Hermes forgets between sessions. Check ~/.hermes/ has files. If empty, re-run the install.
WSL2 can't reach the model server. Enable mirrored networking in WSL settings on Windows 11 22H2+. Or run the model server inside WSL2 instead.
Full tool stack
1Tool Purpose Cost2────────────────────────────────────────────3Hermes Agent agent framework free4 github.com/NousResearch/hermes-agent56LM Studio local model server free7 lmstudio.ai89Qwen 3.6 27B the model free10 via LM Studio or ollama.com1112Stripe payments 2.9% + 30c
Startup cost: $0. Time to first client: one week.
After every delivered report, ask two things. First, a review. Second, one founder they know who might need this.
Founders know founders. By month two referrals replace most of the cold outreach.
The skills folder fills up. Work gets faster. Margin gets better.
Build one report before you have a client. Send it as a sample to 10 people tomorrow.
more setups like this every week. t.me/GipArcAI
