MCP server, with your API key
Pseudonymization MCP server: hand your business processes to AI
Reminders, reconciliations, summaries, onboarding: an agent connected through MCP runs the process from start to finish. With Safe-Doc, it only works on aliases, and the real data is put back at the end, on your computer.
Works with Claude Desktop, Claude Code and any MCP-compatible client. Same key as the API.
safedoc_pseudonymize_documentaged-balance-october.xlsxNames, amounts and contact details are replaced with aliases. The model does not read the original file.agentanalyzes the aged balancePrioritizes the overdue invoices and drafts a reminder for [CLIENT_3], [CLIENT_7]…safedoc_restorereminders.docxThe real names and amounts are put back into the file, on your computer. The model does not see them.youreview and sendA human check before anything goes to a client.
For
What becomes possible
An assistant answers a question; an agent connected through MCP carries out a task. It fetches the documents, compares them, drafts, then submits the result to you. The obstacle was not technical: the data in these processes identifies people, and it could not be sent as is to an AI provider.
- Stable aliases.
[CLIENT_12]stays the same client in the aged balance, the bank statement and the reminder: the agent matches, calculates and drafts correctly. - The model never reads the original. The tool receives the file path, not its content. The model only sees the pseudonymized version.
- The result comes back to you. The real values are put back into a file on your computer, without going back through the model.
What each link in the chain sees
The MCP server does pseudonymization and nothing else. Documents are processed on Safe-Doc servers in Europe, then purged.
- 1
Your files
Accounting exports, contracts, case documents: they stay on your computer. The agent only passes their path to the tool.
- 2
Safe-Doc, in Europe
Detects more than 90 types of data and replaces them with aliases. Purged after processing, with no third-party AI.
- 3
The agent and its model
Claude or another model works on the pseudonymized version: it analyzes, matches, drafts.
- 4
Your computer
The mapping table stays with you. The final result is written with the real names, ready to review.
Four tools, called like any other
The @safedoc/mcp server (npm package) is added to your MCP client. The agent picks the tool it needs at each step.
safedoc_pseudonymize_textPseudonymizes a text extract: more than 90 types of data detected.
safedoc_pseudonymize_documentPseudonymizes a PDF, Word file or spreadsheet from its path. The model never reads the original file.
safedoc_restorePuts the real values back into a file on your computer, using the table. The model does not see them.
safedoc_risk_reportGives a residual risk score and the list of data detected, to check before going further.
Accounting firms: the processes that open up to AI
Most of them rely on data exported from the accounting software. Pseudonymization sits between the export and the agent.
| Process | Data to protect | What the agent does |
|---|---|---|
| Chasing unpaid customer invoices | Customer names, amounts, contact details | Reads the aged balance, prioritizes, drafts the reminders; real names and amounts are put back before sending |
| Chasing missing receipts | Customers, suppliers, bank transactions | Spots transactions without supporting documents and prepares the request for each client |
| Bank reconciliation with the general ledger | Descriptions, counterparties, IBANs | Matches lines on stable aliases and flags discrepancies |
| Summary memo | Figures, executives, employees | Drafts the memo from the pseudonymized statements |
| Monthly summary and management review | Revenue per client, margins, payroll | Comments on the month's indicators for the business owner |
| Preparing the year-end review meeting | Identity, personal and financial situation | Prepares the points to watch and the questions to ask |
| Answering tax questions, with sources | The client's personal situation | Looks up the rules that apply to the situation described in aliases |
| Client onboarding | ID documents, articles of association, company records, IBAN | Checks that the file is complete and extracts the useful information |
| Workforce report | Employees, pay, absences | Computes the aggregates and drafts the report; sick leave is health data |
Lawyers and legal departments
- Case chronology The agent reads the documents and builds the chronology of the facts, without seeing the parties or witnesses.
- Contract review in bulk It lists clauses, deadlines and deviations from the template, with parties and amounts pseudonymized.
- Draft submissions or opinions It prepares a first version from the case file.
- Due diligence questionnaire It fills in the questionnaire from the target's data room.
Human resources
- Annual review summary The agent identifies recurring themes and training needs.
- Employee access request It gathers the relevant documents and spots the third parties to mask.
- Workforce reporting It produces the aggregates and drafts the comments.
- Individual letters It drafts amendments, certificates and replies from the file.
Notaries, consulting and audit, DPOs
- Estate settlement The agent lists heirs and assets from the documents.
- Draft deed It prepares a first draft from the file.
- Data room review It reviews documents in batches and flags points of attention.
- Analytical review It analyzes the general ledger and lists anomalies.
- Access request (DPO) It searches for the person's data and prepares the reply.
- Record of processing It updates the record from supplier contracts.
Good practices
- One path to the data. If another connector gives the agent direct access to the accounting software or the HR system, route that data through Safe-Doc before it reaches the model.
- A human check. Review the result before anything is sent to a client or an employee.
- An up-to-date record. Pseudonymized data is still personal data: list the processing in your record of processing activities.
The risks specific to the protocol and the controls to put in production are covered in the MCP and pseudonymization guide.
FAQ
Which MCP clients does the server work with?
Claude Desktop, Claude Code and any client compatible with the MCP protocol. The server installs as an npm package.
Does the model see my original files?
No. The pseudonymization tool receives the file path, not its content: the model only works on the pseudonymized version.
Where are the real values put back?
In a file on your computer, using the mapping table, which stays on your side. The result does not go back through the model.
Which formats are accepted?
Text, PDF, Word files and spreadsheets.
Which plan do I need?
The MCP server uses the same key as the API, included from the Pro plan. Details are on the pricing page.
What about an agent with no account?
It can have a text anonymized by paying for each call with the x402 protocol, with no sign-up and no key.
Our guides on this topic
GuideMCP and pseudonymization: finally handing your business processes to AIMCP lets you hand an AI agent complete processes (payment reminders, reconciliations, summaries) that were blocked by personal data until now. How to pseudonymize in an MCP flow, the protocol's specific risks and the controls to put in place.Read the guide 
GuidePseudonymizing before an analytical review: the guide for audit and accounting firmsAudit, accounting, general ledgers, financial statements: why and how to pseudonymize data before an analytical review or an AI. Professional secrecy, GDPR, method and evidence for regulators.Read the guide Hand over the processes, keep the data.
A full 10-day trial, no credit card. Your API key works for both the API and the MCP server.