Hermes 0.2 Just Dropped: Everything You Need to Know
Par Miguel Torrez, fondateur de Genial Agency
You will know what changed in Hermes 0.2 (Herald), which features I found most useful, and how to start a local Hermes agent with an OpenRouter key and DeepSeek V4 Flash to try them yourself.
Ce qu'il vous faut
Hermes Agent, desktop app or terminal (hermes-agent.nousresearch.com/docs)
An OpenRouter API key (openrouter.ai)
Optional: an ElevenLabs API key so Hermes answers out loud
Étape par étape
Talk to Hermes with voice mode
Open the Hermes desktop app (voice mode also works in the terminal). Click the small ear icon at the bottom. Set your magic words: Hermes listens and starts a session when you say them. Add an ElevenLabs API key if you want spoken replies. To stop, click end.
Example wake phrase I useHey, Hermes
ElevenLabshttps://try.elevenlabs.io/migueltorrez
Let cited sources check web research
You do not need to turn anything on. When Hermes fetches information from the web, it now runs an extra loop that checks the quotes in the documents it pulled. This reduces hallucinations. It is most useful if your work is heavy on documents and research.
Send data out with outbound webhooks
Give Hermes an endpoint (the address of another app, such as your CRM) and it works out how to send the data there. The automation itself does not need AI at its core: use AI to build it, then let it run.
Customize the desktop app
Drag and drop panels to fit your workspace. Press Command K to open every hidden option. Connect the app to a VPS, your own computer or over SSH. You can see sessions, artifacts, skills, models and the terminal log. If you dislike something, ask Hermes to change it: the app is open source.
Stop clicking approve with Smart Approvals
Hermes reviews the approvals you gave during a day or session, spots the ones you accepted several times, and marks them as always allowed. Ask Hermes to run this once a day, twice a day or every session, or to turn it off entirely.
Turn on steering mid-response
If you send a wrong prompt, you no longer wait for Hermes to finish the mistake. Send a correction and Hermes changes the output mid-response. You can switch this on or off; I keep it on.
Use the new connections, or build one
Hermes now connects to Vercel (website deployment) and Buzz (the Block platform where agents talk to each other). If a tool you need has no connector, ask your Hermes agent to create one, then share it with the community.
Vercelhttps://vercel.com
Start a local Hermes agent cheaply
Go to OpenRouter, get an API key, run Hermes locally and choose DeepSeek V4 Flash as the model. I estimate about $3 a day to run my whole agent, or around $5 for two to three days of trying it out.
OpenRouterhttps://openrouter.ai
Hermes Agent docshttps://hermes-agent.nousresearch.com/docs/
Points d'attention
Voice mode did not work when you connect from your laptop to Hermes on a VPS at the time of recording. The bug was reported and a fix was expected soon, so check before relying on it.
A2A (agent to agent) is still early. I expect 3 to 6 months before it is practical for agents that edit the same files.
Cited sources runs in the background, so do not expect a visible change unless you measure hallucinations.
Every setting (Smart Approvals, steering) can be changed: if a default bothers you, tell Hermes instead of working around it.
