Security & data
Where does my agent live?
On its own server in the region you pick — EU, US, or APAC. Not a shared multi-tenant pool.
Who at Ermes can read my conversations?
No one, by default. Your server is isolated; Ermes staff don’t have standing access. If you ask for support, you can grant time-boxed, read-only access that auto-expires in 24 hours. Every staff session is logged and visible to you.
Do the model providers see my prompts?
Only the prompt and tool inputs needed to answer the current turn. We use zero-retention agreements with providers that support them, so they don’t keep or train on your traffic. With Bring-Your-Own-Key, your prompts go through your account under your zero-retention terms — we never see them.
What encryption is in place?
Disks on your server are encrypted at rest. Everything between you, your messengers, and your agent is TLS in transit. Backups (when present) use keys not shared across customers.
If you have a security incident, when do I find out?
Within 72 hours of confirmation, by email — what happened, what we know so far, what we did, and what you should do.
Trust & control
Does it act on my tools without asking?
No. Sensitive actions stage themselves and ask for one tap. You can require approval on everything.
Can I see what my agent did while I was asleep?
Yes. Every action your agent stages or takes is in a timeline you scroll. Filter by tool, by routine, by day. Nothing happens in the dark.
Can I take a permission back from the agent?
Any permission, any tool, any time. One tap revokes; the agent stops being able to use that tool the same heartbeat.
What’s the "stop the agent" button?
A single switch in your account that pauses every routine and freezes the agent’s active session. Everything queues; nothing executes until you unpause.
What does my agent actually remember about me?
Conversations, decisions you’ve made, people you’ve mentioned, the shape of your week — and the messy stuff: things you said in passing, half-formed ideas, a name you almost forgot. Ask by description, not exact words; it finds what you meant. The open-source agent stacks you can self-host don’t ship that kind of recall out of the box — they keyword-match by default. Memory stays yours: editable line-by-line, exportable as JSON, prunable any time.
Failure modes
What if my agent gets something wrong?
You see what it staged before it sends. You can edit, skip, or undo. Mistakes never become irreversible without your tap.
What if the model hallucinates a fact in a draft?
The agent stages the draft for you. You read it before it goes anywhere. The architecture assumes the model can be wrong — that’s why nothing irreversible runs without a tap.
What happens to scheduled routines if the model has a bad day?
Routines surface their output to you on the messenger you picked. If a routine produces something that doesn’t make sense, you don’t approve it, and you can pause the routine from the same message.
Money
What’s included in the $20?
Your dedicated server, Ermes software, your agent’s brain (via our paid model accounts), one messenger live (Telegram today; others rolling out), unlimited routines, and the support inbox.
Is there really no usage meter?
No tokens, no metered "thinking time", no per-message fee. One price, your agent works as long as you keep talking. (We reserve the right to email anyone who finds a creative way to abuse this — but it’s not a hidden meter.)
What does "bring your own key" actually save me?
If you already pay Anthropic, OpenAI, Mistral, or Gemini, you can route your agent through your own API key. You pay the provider directly for usage and your zero-retention terms apply. Ermes still costs $20/mo for everything else. (Subscription pass-through — ChatGPT Plus, GitHub Copilot — is a separate option in priority beta rollout; see below.)
Can I use my ChatGPT Plus subscription to power my agent, instead of an API key?
Yes, in priority beta access. Where your ChatGPT Plus / Pro / Business plan is eligible, we can route your agent through that subscription — your agent runs on what you already pay for, no second meter, no markup. GitHub Copilot works the same way. Subscriptions are a different reliability profile than API keys (rate limits sized for one human, not an agent), so we keep Ermes’s managed account as the default and let you flip to the subscription once we’ve walked you through it. Email the waitlist if you want this on day one.
What about Claude Pro or Google AI Pro? Same trick?
Claude API keys work today. Anthropic’s consumer subscription terms restrict third-party tools from automating your account, so we explicitly do not offer a Claude Pro pass-through — using a Pro subscription that way would put you offside with Anthropic. Claude Max via Claude Code is documented for advanced developers only. Google Gemini works via API key today; Gemini OAuth (gemini-cli) is on the roadmap behind ChatGPT Plus and Copilot.
What’s the refund and cancellation policy?
Cancel any time from your account. We refund the unused portion of the current month, prorated to the day, and your server is wound down on a schedule you pick. No "cancel via email" friction.
Compliance & legal
GDPR — what are my rights?
Access, correction, deletion, and portability of every piece of data we hold about you. Email [email protected] or use the export and delete buttons in your account. We respond within 30 days; usually much faster.
Do you sell my data, or use it to train models?
No. We don’t sell, we don’t share with advertisers, we don’t train models on your data, and we hold contracts with our model providers that say the same. Sub-processors are listed publicly.
Are minors allowed?
Ermes is for adults (16+ in the EU, 18+ elsewhere) per our Terms. We don’t knowingly accept accounts from younger users.
Comparisons
What is the easy, safe way to get an AI agent that is yours?
A chatbot tab is the fastest AI surface (about a minute) but it is not an agent. A multi-tenant AI assistant is the quickest agent-shaped product, on shared infrastructure. Ermes is the easy, safe way to get an AI agent that is yours — your own server, your own memory, consented context, and approval before anything sensitive is sent. You pick a name and a region; we provision a dedicated agent in your name and you can disconnect or delete in one tap.
How is this different from ChatGPT?
A chatbot writes the answer; your agent finishes the task — in the messenger you already keep open.
How is this different from a typical multi-tenant AI assistant?
Most AI assistants run thousands of customers through one shared database, one shared model account, and one shared memory layer. It’s cheap and it scales — but it means another customer’s bug, another customer’s prompt injection, or a vendor support engineer with the wrong query can reach into your conversation. Ermes doesn’t do that. Each account gets its own server, its own model account, its own memory store. There is no "everyone’s data" table for something to leak from, because there is no everyone — there’s you.
Portability
Can I export my agent’s memory?
Yes. Prune any item, export everything as JSON, or wipe the agent — from your settings.
Can I migrate to a different provider later?
Memory and routines export as portable JSON. We don’t lock in your data. (We do believe the Ermes shape — dedicated server, $20 flat — is the right one; but if you decide it isn’t, you leave with everything.)
What happens to my agent if Ermes shuts down?
First, the honest part: we’re a small company. The most important promise we can make is that if we ever wind down, you don’t get caught flat. At least 30 days notice before the lights go off — and we’d aim for 90+. A final export of your agent’s memory, routines, and conversation history as portable JSON, valid for at least 30 days after shutdown. Bring-Your-Own-Key continuity: if you’re already on BYOK, your model account is yours — it keeps working with whatever tool you point it at. If you’re on Ermes’s paid model accounts, we’ll guide you through moving to your own key during the wind-down so the conversation doesn’t end the day we do. The Ermes software itself is ours, and it stops with us — we’re not promising you’ll keep running an agent named after you forever. We’re promising your data, your habits, and your model continuity are not held hostage to our cap table.
Beta state
What’s live versus coming?
Telegram messenger, dedicated server provisioning, routines, the permissions UI, and BYOK are live. WhatsApp and iMessage are in private beta. Signal, Discord, and Slack are next.
Should I expect things to break?
Occasionally. We’re in private beta — we ship, we monitor, sometimes a routine misfires or a messenger reconnect is rough. We post incidents publicly and credit you for downtime that affected your account.