AI-assisted access to every client site you manage — however it's hosted.

It is where the agent runs, and what that costs you in control. Remote Agent is the default: the agent thinks inside AlmaSEO and sends one command at a time down the SSH connection, so the platform can inspect each one, hold anything that changes the site for your approval, back files up before they are edited, and strip secrets out of what the model sees. That is the mode all the safety guarantees on this page describe.

Direct Agent installs Claude Code onto the server itself — one click on a saved connection, no terminal — and drives it from the same window. The agent is on the box, so it reads and edits at disk speed instead of a round trip per file, and it can follow a long investigation to the end rather than working to a step budget per message. It runs on the same Anthropic API key you already gave AlmaSEO; there is nothing extra to buy.

The trade is the gate. In Direct mode AlmaSEO is not standing between the agent and the server, so there are no approval cards, no file backups, and no rollback — it edits when it decides to. It is off until you install it, per connection, and one click in the session header switches back. Start in Remote, and reach for Direct on the jobs where you would hand a developer the same access.

Yes — safe access is the foundation this feature is built on, not an afterthought. Because AlmaSEO connects directly to your server to diagnose issues from the inside, we designed it so that nothing can change or damage your site without your explicit approval.

Here's how it works. AlmaSEO reads and investigates freely, but the moment it wants to run anything that could modify a file, alter your database, or send data off your server, it stops and asks you first. You see exactly what it wants to run and why, and it doesn't proceed until you click approve. You're always the final decision-maker — the AI can recommend, but it can't act on anything sensitive on its own.

We also assume the internet is a hostile place, and we built accordingly. Your login credentials are encrypted, and your server's identity is verified on every connection, so AlmaSEO won't be fooled into connecting to an impostor. If we ever read something sensitive like your site's configuration file, the AI can use it to help you, but it's automatically hidden from our activity logs rather than stored in plain text. And because AI assistants can be manipulated by malicious content, AlmaSEO is specifically instructed to treat anything it finds on a server as untrusted — it will never follow hidden instructions buried in a file or a log.

Finally, these protections have been put through repeated adversarial hardening reviews built to break them, and each command AlmaSEO runs is permanently logged, so you have a complete, reviewable record of everything that happened.

All of that describes Remote Agent mode, which is how every session opens. There is a second mode, Direct Agent, that installs Claude Code onto the server and lets it work there without the approval step — faster and deeper, and correspondingly less supervised. It is opt-in per connection and clearly labelled in the session header, so you are never in it without having put yourself there.

The AI doesn't enforce them, so it can't skip them. Every command it wants to run is inspected by the platform before it executes, and the platform decides whether it's allowed to run on its own or has to stop and ask you. Read-only commands go through. Anything that modifies a file, touches the database, or reaches off the server comes back to you as an approval card.

That distinction matters. A lot of AI tools ask the model to police itself — the safety rule lives in the prompt, and a clever enough input can talk it out of them. Here, the gate sits outside the model entirely. The AI can be wrong, or confused, or fed a malicious instruction hidden in a log file, and the answer is the same: it still doesn't get to run a destructive command without your click.

On top of that, approval cards tell you whether the change is reversible before you approve it, files are backed up before they're modified, and anything that goes wrong can be rolled back.

One mode is deliberately outside all of this, and we would rather say so than let you find out. Direct Agent mode installs Claude Code onto the server and lets it work there directly — no round trip through AlmaSEO, so nothing for AlmaSEO to gate. It is off until you install it on that specific connection, the session header shows which mode you are in, and one click puts you back in the gated one. Everything above applies to Remote Agent mode, which is where every session starts.