Claudia
Claudia is the assistant built into the platform. The idea behind her is the idea behind everything else here: nothing is programmed, it is configured — and configuring can be a conversation.
Back office → System → Claudia.
The one thing to understand first
Claudia does not have her own way in.
Every action she takes goes through the same domain service the back office uses, as a specific person, with that person's permissions, in that person's organization, and it lands in the activity trail under her signature.
She cannot exceed the operator
The user she acts as is fixed by the server when her tool server starts — it is never something the model can pass as an argument. There is no tool call that can change who she is, because who she is was decided before the conversation began.
That single decision is what makes the rest safe: an assistant with more access than its user is not a tool, it is a way around the permission system.
And in the organization they have open. When you ask her for "the sizes" or to "delete that register", she answers about the firm in your toolbar — never about another firm that happens to use the same name for something. Two firms may each have a Size; they are two different registers, and she works on yours.
If she says she cannot find it, check which firm you have open
That is usually the answer. A register belonging to another firm is not hidden from her out of caution — it is genuinely not hers to see while you are working in this one.
What she can do
Roughly everything the back office can, because her tools are the back office's services:
Structure
Content types, fields, field groups, displays, listings, taxonomy, menus, media types, forms, addresses, meta tags, workflows.
Business
Partners and CRM, documents, inventory, projects, helpdesk, HR, mail, campaigns, registers, accounting configuration.
She can also administer herself — her settings, her briefing, her model and budget, which tools are switched on, her preset prompts, and the seats. Those tools carry the same permission the Claudia admin pages do, and they are write-only for secrets: no tool ever returns an API key or a seat token.
Tool parity
Every agenda exposed to Claudia has to be exposed completely: if she can create a thing, she can also change it and delete it.
Why a half-exposed agenda is worse than none
A capability that can create but not delete strands whoever used it halfway — they now have records only a human can clean up, made by an assistant they were told could manage them. A test enforces the parity, so an agenda cannot be added to her surface in half.
Refusals are answers, not crashes
When a service refuses her — validation, permissions, a boundary — she receives a structured refusal naming the field and the reason, and she corrects herself. She never sees a traceback, and a refusal is never converted into something that looks like success.
This matters more than it sounds: an assistant that cannot tell "refused" from "failed" will retry a permission error forever.
Settings
| Setting | Decides |
|---|---|
| Edition | Free or paid |
| Signature | The name her actions are attributed to in the trail |
| Persona, greeting, extra instructions | How she talks and what she always keeps in mind |
| Default model | Which model answers |
| Monthly budget | The ceiling, with a warning threshold |
| Disabled tools | The agendas she may not touch at all |
| API fallback | Off by default, and a deliberate choice |
The tool surface is yours to narrow
Disabled tools is the switch worth knowing. If an agenda should never be
touched by an assistant in your firm — accounting, say, or mail — switch it
off and it disappears from her surface entirely. She cannot ask for it back.
The morning briefing
Claudia can write a briefing on a schedule: what arrived in the mail, which invoices are due, what is still a draft, what the calendar holds. You choose the days, the time, the model, and which of those four she includes.
It runs as a named user, so the briefing sees exactly what that person sees — a briefing that could read more than its reader would be a leak with a friendly tone.
Preset prompts
A prompt you keep re-typing becomes a preset: a name, the prompt, optionally its own model, and optionally a schedule. Scheduled presets also run as a named user.
Memory
Claudia keeps a memory per organization — the standing facts about how your firm works, so you do not repeat them every morning. It belongs to the organization, like everything else, and it does not cross to another.
Cost
Every conversation records its tokens and its cost. The monthly budget is a ceiling with a warning at a percentage you set. Licensing is per seat, one to one with the people using her.
Photos in the chat
Show her a photo — a receipt, a label, a damaged delivery, a screenshot. Use the 📎 button (on a phone it offers the camera), paste it with Ctrl+V, or drop it onto the panel; up to six per message. Large photos are shrunk in the browser first, so sending is quick.
Each photo is filed in Files as your own private file in the firm you have open — nobody else sees it, not even a superuser — and the conversation shows it again when you come back to it.
Licences: Claude or ChatGPT
System → Claudia → Licences. A seat gives one person Claudia on one AI account — and the account can be Claude (Anthropic) or ChatGPT (OpenAI). Pick the provider on the seat; everything else stays the same: the same panel, the same tools, the same language, the same history.
Nobody signs in on the server. You activate a seat by pasting the sign-in you made on your own computer:
| Provider | On your computer | Paste into the seat |
|---|---|---|
| Claude | claude setup-token, sign in with the person's Claude account |
the printed token |
| ChatGPT | npm install -g @openai/codex, then codex login with the person's ChatGPT account |
the whole content of ~/.codex/auth.json (Windows: %USERPROFILE%\.codex\auth.json) |
The sign-in is stored encrypted and write-only — the form only ever says that one is stored. A ChatGPT sign-in renews itself on the server between turns; you paste again only if the person revokes it or you change the account. Changing a seat's provider deletes the old provider's sign-in.
Test on the seat runs one short turn on that seat, as its person, and tells you the provider, the model and the time — or the provider's reason for failing. The test is logged like any other turn.
On ChatGPT she is still a manager, not a programmer
On either provider Claudia only talks and works through the platform's own tools, as the signed-in person. She has no files, no commands, no web browser and no code — the ChatGPT agent's own abilities of that kind are switched off, and its sandbox refuses any attempt to write.
Models. Model and limits keeps a default and a briefing model for Claude and, separately, for ChatGPT (empty = the account's default).
The API fallback — for people without a seat, off by default — runs on the provider you choose there, with that provider's key (an Anthropic key or an OpenAI key). The monthly budget applies to it on either provider. The run log shows the provider of every turn.