Your first real conversation
Three small tasks that prove Hermes is actually doing things, not just talking.
This is the moment it clicks. You're going to give Hermes three small jobs — each one touches a different part of the agent — and watch it actually work. Ten minutes, and you'll know why an agent is not a chatbot.
Start a session
hermes
You're now in an interactive chat. Type your messages, press enter. Simple so far.
Task 1: Ask it a real question
Start with something that needs the web:
What's the weather like in Tokyo right now?
Watch closely. You'll see Hermes decide to use the web search tool, run the search, read the results, and summarize them back to you. It doesn't just hallucinate an answer from its training data — it fetches real information.
If the web toolset isn't enabled, Hermes will tell you. Turn it on with /tools enable web and ask again.
Checkpoint — You should see Hermes run a tool call (usually labeled "search" or "web") before replying. If the reply mentions tool use, you're good.
Task 2: Make it read a file
Create a small test file so there's something to look at:
echo "Hermes is an AI agent that runs in your terminal. It can use tools, connect to messaging platforms, and remember things between sessions." > /tmp/test.txt
Back in your Hermes session, ask:
Read /tmp/test.txt and tell me what it says in one sentence.
Hermes will call the file-read tool, open the file, and summarize the contents. This is the same mechanism it uses when you ask it to debug a codebase — it reads the files it needs.
Task 3: Run a shell command
Now the scary one. Ask:
Run `ls ~` and tell me how many files and directories are in my home folder.
Hermes will ask for your permission the first time it tries to run a shell command. That's intentional — the security model defaults to asking before it touches your system. Approve it, and watch it run the command, count the output, and give you a number.
If you don't want the approval prompt every single time, you can switch to smart approval mode later. We'll skip that for now.
Checkpoint — If the three tasks all worked, you've just used the three biggest tool categories in Hermes: web, file, and terminal. Everything else is a variation on these patterns.
When things go sideways
It gives you an answer without using tools. The provider might be trying to answer from memory instead of searching. Ask more explicitly: "Use the web to find out" or "Read the file at this exact path." Or check /tools to see which toolsets are enabled.
Permission prompt keeps appearing. That's the security system working. You can review and approve each command. For terminal-heavy workflows, you'll want to read up on approval modes later.
The model is slow or responses time out. Try switching to a faster model with /model. For agent work, smaller-faster is often better than bigger-smarter — the agent can recover from small mistakes but not from timeouts.
Wrap up
Exit the session when you're done:
/quit
If everything worked, Hermes has now shown you three things it can do that a chatbot cannot: it searched the live web, read a file from your disk, and ran a command in your shell.
Next, we're going to learn the five slash commands you'll actually use every day.