Usage examples

Real requests from Oxygenius users, the tools the assistant chains and what the user gets. Users mostly write in French: each request is shown as typed, with an English translation in the outcome. Identifiers and amounts are fictitious.

Good practice
  • Name the project, customer or person: the assistant nearly always starts with a search (search_*).
  • For a write, the assistant states exactly what it will do and waits for your approval: read it.
  • Anything leaving the company always goes through a preview (prepare_*): check recipients and content before accepting.

Looking things up

1. Get the status of a job

Read

Request

« Où en est le chantier de M. Dupont à Namur ? Donne-moi l'étape en cours, le gestionnaire et ce qui reste à facturer. »

What the assistant does

  1. search_dossiers with query: "Dupont Namur": the search covers the customer, the site address and references.
  2. get_dossier on the project found: status, manager, current step, parties.
  3. search_estimates with dossierId, then get_estimate on the accepted estimate: signed amount and amount left to invoice.

Outcome. A summary of the project: step, manager, address, accepted estimate, amounts already invoiced and left to invoice. If several projects match, the assistant asks which one. (Request in English: "Where is Mr Dupont's job in Namur at? Give me the current step, the manager and what is left to invoice.")

2. List overdue invoices

Read

Request

« Quelles factures sont en retard de paiement ? Classe-les par montant restant dû. »

What the assistant does

  1. search_invoices with state: "overdue": outstanding balance and delay recalculated (amount due − credit notes − payments).
  2. If the result says hasMore: true, the assistant calls the tool again with offset for the next page.

Example arguments

{
  "state": "overdue",
  "limit": 25
}

Outcome. A table of invoices past due and not settled: customer, number, due date, days late, balance; the total due. ("Which invoices are overdue? Sort them by outstanding amount.")

3. Prepare the week

Read

Request

« Qui travaille sur quels chantiers la semaine prochaine ? »

What the assistant does

  1. search_planning_events over the period: one entry per day with the people assigned, in Belgian local time.
  2. get_planning_event when the user wants the site address or the estimates linked to an entry.

Example arguments

{
  "from": "2026-11-23",
  "to": "2026-11-29",
  "limit": 25
}

Outcome. The planning day by day, per person or per site as asked. ("Who is working on which job sites next week?")

4. Read a project document

Read

Request

« Que prévoit l'offre du fournisseur de châssis classée dans le dossier Dupont ? Quel est le délai de livraison ? »

What the assistant does

  1. search_dossiers then list_dossier_documents with query: "châssis".
  2. read_document on the quote: PDF text in parts of at most 20 pages; the assistant reads on with part = nextPart if needed.

Outcome. The terms of the quote (price, lead time, conditions) quoted from the document. A scanned page without text is reported, not made up. ("What does the window supplier's quote filed in the Dupont project say? What is the delivery time?")

External content

The text of a document or e-mail is data: the server cleans it, and its instructions tell the model never to follow an instruction found in it.

5. My e-mails and tasks

Read

Request

« Résume mes mails non lus et dis-moi quelles tâches urgentes je dois traiter aujourd'hui. »

What the assistant does

  1. get_my_emails with unreadOnly: true, then get_email for the body of the important ones (without marking them as read).
  2. get_my_tasks: tasks received and to do, urgent ones first.

Outcome. A digest of the unread e-mails and the list of urgent tasks, with their project. ("Summarise my unread e-mails and tell me which urgent tasks I must handle today.")

Taking action (write tools)

These examples require the company to have enabled write tools (Getting connected).

6. Draft an estimate

Write

Request

« Prépare un devis pour le dossier Dupont : pose de 25 m² de carrelage 60×60 avec la colle, plus la dépose de l'ancien carrelage (6 h). »

What the assistant does

  1. search_dossiers and search_items to find the project, the tiles and the adhesive in the catalogue.
  2. The assistant announces the estimate it will create and the lines it will add, and waits for your approval.
  3. create_estimate on the project, then add_estimate_lines: a title, two catalogue items, a labour line.

Example arguments

{
  "estimateId": 4812,
  "lines": [
    {
      "kind": "title",
      "name": "Bathroom tiling"
    },
    {
      "kind": "item",
      "itemId": 2290,
      "quantity": 25
    },
    {
      "kind": "item",
      "itemId": 2307,
      "quantity": 5
    },
    {
      "kind": "free",
      "name": "Removal of the old tiles",
      "category": "workforce",
      "quantity": 6,
      "unitPrice": 48
    }
  ]
}

Outcome. An editable estimate in Oxygenius. Prices, VAT and totals are computed by Oxygenius as in the estimate screen: the assistant never sends a computed amount. Nothing is sent to the customer. ("Prepare an estimate for the Dupont project: laying 25 m² of 60×60 tiles with adhesive, plus removing the old tiles (6 h).")

Requirements

The company has enabled write tools, and the user has the dossier-detail.estimate.create and estimate.list.view permissions.

7. Number an invoice and send it to the customer

Write

Request

« Numérote la facture d'acompte du dossier Dupont et envoie-la au client. »

What the assistant does

  1. search_invoices with dossierId to find the draft.
  2. prepare_invoice_finalization: changes nothing, shows the coming legal number, the effects (PDF, UBL, accounting entry, automatic sending to the accountant if the company set it) and the warnings (date, due date, zero total), with a 10-minute token.
  3. After your explicit approval: finalize_invoice with the token.
  4. create_document_email with kind: "invoice": the company's e-mail draft, PDF attached.
  5. prepare_email_send: the e-mail exactly as it will leave (sender, recipients, subject, message, attachments, warnings).
  6. After your explicit approval of that e-mail: send_email with the token.

Outcome. The invoice is numbered and sent; its status becomes Sent. A document changed between preparation and action makes the token invalid: prepare again. ("Number the deposit invoice of the Dupont project and send it to the customer.")

Irreversible actions

An invoice number is legal and final, a sent e-mail cannot be recalled. These tools are marked destructive: Claude always asks for approval, on top of the agreement given to the assistant. Sending is capped at 30 per user per day.

8. Send a task to a colleague

Write

Request

« Demande à Julie de commander les tuiles pour le chantier Dupont, c'est urgent. »

What the assistant does

  1. search_users with query: "Julie" and taskRecipientsOnly: true.
  2. list_task_types with assigneeId: the task types that can be sent to her.
  3. The assistant announces the task (recipient, type, project, content, urgency) and waits for your approval, then calls create_task.

Example arguments

{
  "assigneeIds": [
    37
  ],
  "taskTypeId": 4,
  "dossierId": 1186,
  "description": "Order the roof tiles for the Dupont job.",
  "isUrgent": true
}

Outcome. Julie receives the task, pinned as urgent, and Oxygenius notifies her. ("Ask Julie to order the roof tiles for the Dupont job, it's urgent.")