Data sent to Anthropic
NomaUBL uses Anthropic's models in two ways, and the path of the data is not the same:
- the AI assistant of the application — the server calls Anthropic's API with the API key of the instance;
- Claude connected through MCP — NomaUBL calls nobody: it answers the tool calls of the user's Claude client, and that client talks to Anthropic under the user's own Claude account.
Nothing is sent while nobody asks a question: no background send, no telemetry, no other AI provider.
At a glance
The AI assistant
Each question is sent to api.anthropic.com with the API Key of Settings → Global → AI. A request holds:
| Sent | Content |
|---|---|
| Instructions | The built-in NomaUBL prompt, or the System Prompt of the instance in its place; the status catalogue; the names of the playbooks. |
| Page context | The current page and its filter, the language, and the open invoice: document, type, company, number, status, customer. 4,000 characters at most. |
| Conversation | The questions and the answers of the conversation in progress — the last 40 messages, kept in the user's browser. |
| Tool results | What each tool read to answer, as it read it: invoices (customer name, amounts, VAT, status, messages), lines, lifecycle, validation errors, period figures, dashboards, e-reporting runs, e-directory data, lines of the server log, sections of the implementation report, settings with secrets masked. A result is cut at 60,000 characters; a list holds 50 invoices at most, the server log 300 lines. |
| Attached file | An XML as text, up to 150,000 characters; a PDF as it is, up to 10 MB. The file is sent again with each later question of the same conversation. |
Two other screens call the same API:
| Screen | Sent |
|---|---|
| AI mapping of the XSL editor | The paths of the sample with the first value of each (120 characters at most), the text rows of a PDF, the list of variables — and the PDF itself when the sample is one. |
| Documentation in Sitemap mode | Nothing more from NomaUBL: Anthropic fetches the pages of the allowed documentation domains itself. |
Business data is sent as the tools read it: no customer name, amount or e-mail address is replaced or hidden. Only the secrets are masked — see below.
Claude through MCP
| Step | Who talks to whom |
|---|---|
| The user asks Claude a question. | The Claude client ↔ Anthropic, under the user's Claude account. |
| Claude calls a NomaUBL tool. | The client → the instance, with the user's personal token. |
| NomaUBL answers. | The instance → the client: the result of that tool, 60,000 characters at most. |
| Claude writes its answer. | The client → Anthropic: the conversation, with the results it received. |
NomaUBL makes no call to Anthropic here, and the API key of the instance is not used. What a tool returned leaves the instance towards the user's Claude client, and from there to Anthropic on the terms of the user's Claude plan — not on those of the instance's API key.
Where the call comes from depends on the client:
| Client | The instance is called from |
|---|---|
| Claude Desktop with the extension, Claude Code | The user's computer: an internal address or a VPN is enough. |
| A claude.ai connector | Anthropic's servers: the instance must be reachable from the internet. |
The Claude Desktop extension runs a small relay on the user's computer. It reads one folder only — the Samples folder chosen in its settings — and sends NomaUBL a file of that folder only when the user asks Claude to work on it.
What is never sent
| Never sent | How |
|---|---|
| Passwords, tokens, API keys, client secrets, the licence key | A setting whose name marks it as a secret is given as ***; the licence is described by its state, never by its key. |
| The personal tokens | The server keeps their fingerprint only. |
| A file nobody attached or asked for | The assistant reads the attached file; over MCP, Claude reads the user's own samples. |
| Anything a switched-off tool would read | Custom Tools at N: the assistant keeps the documentation only. MCP endpoint at N: no client can connect. |
| Changes | The assistant only reads. Over MCP, a change needs Allow changes over MCP, and secrets cannot be set that way. |
What stays on the server
| Kept | Where, how long |
|---|---|
| The conversation of the assistant | In the user's browser (the last 40 messages); not on the server. |
| Sample files and files attached in the assistant | On the server, per user, 7 days. |
| AI usage | For each turn of the assistant, of the auto-map and of Claude (MCP): who, when, the model, the tokens, and each tool's name and input summary — for 90 days by default; never the question, the answer or what a tool returned. See AI usage. |
| MCP calls in the server log | One line per failed call and one per change, with the user, the tool and the beginning of its arguments and of its result. A successful read leaves no line. |
What decides
| Switch | Where | Effect |
|---|---|---|
| API Key | Settings → Global → AI | No key: no assistant, nothing sent. |
| Custom Tools | Settings → Global → AI | N: the assistant answers from the documentation, without reading the data of the instance. |
| MCP endpoint | Settings → Global → AI → MCP server | N (default): no Claude client can connect. |
| Allow changes over MCP | Same group | N (default): read only. |
| MCP access | Roles → Access → Features | Which users may connect Claude; no role by default. |
| Explorer page, Settings access | Roles | Which tool groups a user's Claude reaches. |
| Server log and IT dashboard pages, Settings access | Roles | Who may have the server log, the IT dashboard and the implementation report read, in the assistant and from Claude. |
| Companies and row filters | Roles → Data scope | The invoices and e-reporting runs a question can reach. |
To know before opening it widely
- The tools read what the screens show. A customer name, an amount, an e-mail address of the users list in the implementation report go to the model when a tool reads them.
- Invoices and e-reporting runs follow the role. The assistant and Claude read those of the user's companies and row filters, as the lists of the application do.
- The server log, the IT dashboard and the implementation report are those of the whole instance. A tool gives them only to a role that has the matching page — Server log, IT dashboard — or, for the implementation report, Settings access.
- What the user types or attaches is theirs to check. A file attached in the assistant is sent as it is.
- The use Anthropic makes of what it receives depends on the contract of the instance's API key for the assistant, and on the user's Claude plan for MCP.