AI · 1 / 23
Ortha CMS AI
An assistant inside the admin that works with your permissions, asks before every change, and never publishes.
The summary can lose half its words, and the entry has no SEO description yet. I have drafts for both — saving them needs your go-ahead.
content_propose_update {
"id": "e41f0a2b-…",
"values": {
"summary": "Sleeper routes that closed a decade ago are running again across Europe.",
"seoDescription": "Night trains are back: which sleeper routes reopened, and why travellers are choosing them."
}
} Nothing has happened yet — it is waiting for you.
Understood. Nothing was changed.
Enter to send, Shift+Enter for a new line
How it works
- 01
Open it from any page with ⌘J and ask in plain language.
- 02
Reads run straight away. Every write stops at a prompt until you allow it.
- 03
A saved change lands as an ordinary version. A person presses Publish.
- Models
- Claude, or any OpenAI-compatible endpoint
- Permissions
- Exactly the signed-in user’s
- Publishing
- Never — it has no tool for it
In depth · For teams who want help without exposure
An assistant you can give the whole team
It runs inside the same permission system as every other request, so handing it to everyone is a settings decision.
It can only do what you can do
There is no service account. Every run acts as the person who started it — their role, their workspaces, nothing more.
Under the hood
Permissions are resolved per run. A tool you cannot call is never offered to the model, and every call is checked again when it runs.
It asks before every change
Reads run straight away. Before a write it asks: Allow once, Allow for this chat, or Don’t allow. A refusal changes nothing.
Under the hood
“Allow for this chat” ends with the conversation; there is no “always”. Every call it skips the prompt for is still checked against live permissions.
A short exchange in the assistant panel, ending in the permission prompt with its three choices: Allow once, Allow for this chat and Don’t allow.
What it does in the editor
- Finds things
- Searches entries, relations and revisions — “what changed here?” works in words.
- Drafts and edits
- Its changes are validated, versioned and audited like yours.
- Translates
- Drafts a translation into a configured locale, one entry or a batch.
- Writes alt text
- For the images that have none.
- Reports what is wrong
- Reads the findings your alarm rules raised, each linked to its entry.
- Proposes who may read something
- Puts an existing audience on a record, as a proposal you approve.
Bring your own model
Claude through a native adapter, or any OpenAI-compatible endpoint — Ollama, vLLM, llama.cpp, LM Studio, LiteLLM, OpenRouter, Azure. A model on your own hardware keeps content in the building.
Under the hood
The feature is off until an operator turns it on and picks a provider.
Questions
Can it publish?
No. Publishing is not a tool at any role. It can prepare a draft; a person presses Publish.
Can I undo what it did?
Yes. Every change writes a version, and the activity log names you as the actor.
Can I cap what it spends?
Yes. Runs are bounded by steps, time and tokens, and operators can set daily token quotas per person and per workspace.