Hermes on iPhone and iPad
The agent you run on your own computer, in your pocket. Watch every command it runs, answer the approvals it asks for, and pick up any conversation where you left it.
Talaria is a native iPhone and iPad client for Hermes Agent — the open-source agent you
run on your own computer. It is not a chatbot with a server behind it. It talks to your
machine, runs your models, and holds your credentials on your device.
REQUIRES HERMES AGENT
You need Hermes Agent running on a computer you control: github.com/NousResearch/hermes-agent
Talaria connects to it over your own network. There is no Talaria account and no service
in the middle. If you do not run Hermes, this app has nothing to connect to — open it and
tap "Look around with sample data" to see how it works first.
EVERY CONVERSATION, EVERYWHERE
Sessions you started in the CLI, in the desktop app, or through a chat platform all show
up in one list, with previews and the branch each was working on. Pick one up on your
phone exactly where you left it.
WATCH IT WORK
Replies stream in as they are written. Every tool run is there to read — the command, the
output, the file it touched — instead of a spinner and a summary. Charts and images the
agent generates appear inline.
YOU STAY IN THE LOOP
When the agent is about to do something destructive, it stops and asks. Approvals,
clarifying questions, sudo requests and credential prompts arrive on your phone, and
nothing runs until you answer.
TALK TO IT
Hold the microphone and speak. Recognition happens on your phone whenever iOS has the
offline language model installed for your language; where it does not, Apple's speech
service transcribes the audio. Choose "On the server" for either direction in Settings to
send it to your own gateway instead.
SCHEDULED TASKS
See what your agent is set to do and when, read what happened on past runs, and edit the
schedule, the model and the prompt from your phone.
SKILLS, TOOLS AND MEMORY
Browse installed skills and turn them on or off. See connected MCP servers and the tools
they expose. Read the agent's memory. Browse the working directory.
WHAT IT COSTS TO RUN
Token counts and spend per model, so you can see what a conversation actually cost.
CHOOSE THE MODEL PER CONVERSATION
Set the model and the thinking level before you send the first message, or change either
mid-conversation. Eight thinking levels, from off to ultra.
HOW IT CONNECTS
Over HTTPS to a gateway with a valid certificate, or with a prompt to inspect and trust
a self-signed one — the same way an SSH client asks about a host key. Plain HTTP is
allowed only to addresses that cannot leave a private network: Tailscale, your LAN, or
localhost over a forwarded port. Credentials live in the iOS keychain.
PRIVACY
No account. No analytics. No crash reporting. No server operated by us, because there
isn't one. Your data goes to the gateway you point it at — and, if you dictate on a device
without the offline language model, the audio goes to Apple for transcription. Nowhere
else.
Talaria is not affiliated with Nous Research.