Hermes: Build a Self-Learning Agent with 7 Workflows

Most agent frameworks hand you a toy that stops being useful the moment you close the laptop. Hermes is different in one specific way: it learns. Every ~15 tool calls, it reviews what happened and writes a skill file that makes the next run better.

This is the plain-English guide to getting value out of it in a week.

What Hermes actually is

Hermes is a personal automation agent. It lives on a server (or your laptop), runs 24/7, and reaches you through messaging — Telegram, Discord, Slack, WhatsApp, Signal, Email, and a dozen more. It can also run via CLI.

Install is two minutes:

# download the install script
curl ...
# run
hermes
# pick a model
hermes model
# connect Telegram
hermes gateway setup

Drop it on a cheap VPS. Message it when needed. It remembers what's already in flight.

The learning loop is the whole point

Every ~15 tool calls Hermes pauses, reviews what worked and what failed, and writes a skill — a markdown file in ~/.hermes/skills/ that codifies the process. These files aren't hidden; open them, edit them, delete the bad ones.

The practical difference shows up over weeks:

• Day 1. Generic output.
• Day 30. More accurate, more relevant, formatted the way you actually want.

Claude Code's memory stores facts about your preferences. Hermes stores executable procedures. Not "you like bullet points" but the full "research → filter → format" pipeline that produces the bullet points your way.

Hermes vs. Claude Code vs. OpenClaw

• Claude Code lives in your repo. Reads code, writes code, runs tests, commits. One of the best code agents. Doesn't live on a server.
• OpenClaw lives on your server. Messaging, scheduling, tool access. No learning loop.
• Hermes is OpenClaw plus the learning loop. Coming from OpenClaw? One command migrates persona, memory, skills, API keys, and gateways: hermes claw migrate.

7 workflows worth building

1. Morning briefings that learn

Connect Hermes to Telegram with hermes gateway setup. Tell it to monitor email, calendar, and a few topics. Schedule it for 7am.

The summary itself is fine on day one. The value arrives two weeks later, when Hermes has learned which senders you reply to, which meetings you prep for, and which topics pull follow-ups out of you. The skill file updates itself.

2. Website monitoring without manual review

Hermes ships with Camoufox, a stealth browser that doesn't get flagged like standard headless automation. Pair it with Firecrawl for structured extraction and you have a monitoring pipeline that sites don't block.

Point it at competitor pricing, job boards, news feeds, product catalogs. Hermes tracks what changed versus yesterday. No more opening ten tabs every morning — you get a diff.

3. One agent, whole company

One fintech founder tried five specialized agents — marketing, sales, engineering, community, briefings. It collapsed in 48 hours: no shared context, duplicated skills, inconsistent brand voice.

He rebuilt it as a single Hermes instance. Claude Code handles code. One Hermes runs marketing, outreach, community, and briefings — the marketing context is present during outreach because it's the same agent. Shared memory produces compounding context. Five disconnected tools don't.

4. A knowledge base that gets smarter

Hermes ships a built-in skill based on Karpathy's LLM-Wiki pattern. Tell it to create a wiki, point it at sources, and it organizes everything into linked markdown — summaries, entity pages, concept pages, comparisons. Three layers:

• Raw sources are never modified.
• Wiki pages are written and maintained by the agent.
• A schema file enforces the rules.

Add a new source and Hermes doesn't just drop it in a folder. It checks existing pages, updates what changed, adds cross-links, flags contradictions. Browse 643 community skills with /skills.

5. Experiments that optimize themselves

Give Hermes a metric — email open rate, landing page conversion, response time — and tell it to make small changes and keep what works. Autoresearch in a loop.

The learning loop means it doesn't test randomly. It gets better at predicting which changes move the metric. One user wired Hermes to a broker API and ran four automated trading strategies on a live account. Another automated token operations on Solana. Production, real money.

6. Plug in your existing MCP servers

Hermes v0.8.0 added native MCP client support. Every MCP server you already built or configured for Claude Code works with Hermes — Google Workspace, databases, custom APIs. Hermes discovers them. No rebuild, no reconfigure.

7. Pick the right model or lose a weekend

The #1 reason Hermes setups "feel broken" is the wrong model. A model that can't handle tool calls will hallucinate tools that don't exist — people blame the framework, but the framework is fine.

For local experimentation, Gemma 4 26B via Ollama is the best open-source option. For the workflows above, use a frontier-model API. Switch with:

hermes model

Nous Portal, OpenRouter's 200+ models, OpenAI, or your own endpoint. No code changes, no lock-in. Run hermes doctor if something feels off — it diagnoses config issues before you burn hours guessing.

The bottom line

Hermes is the closest thing to a personal agent that actually improves with use. Install it, connect Telegram, give it one recurring task, and let it run for two weeks before you judge it. The version at day 30 is the one worth evaluating — not the empty one you spin up on day one.

If you want to pair it with a code agent, see the Claude tools guide or the managed agents playbook.