SiberMate

Remote MCP

MCP data and privacy.

When you use SiberMate MCP, tool results go to your AI client and from there to the model provider you chose. This page covers what leaves SiberMate, what is masked before it leaves, and what SiberMate MCP keeps.

Markdown version Interactive MCP guide

Emails and names are masked before they reach the AI

  • Email addresses are shown as b***#1rloay@acme.co.id. The company domain stays; the person's part does not.
  • The 6 characters after # are a tag computed from the whole address. The same email always gets the same tag, and two people who share an initial and a domain get different ones. The tag shows no letters of the address, but because it never changes, the AI provider can tell that two results are about the same person, and someone who already holds a list of candidate addresses could check whether one of them is in your tenant.
  • Person names are shown as B*** S***. This covers learners, admin users, group managers, policy owners, and learner rows in reports.
  • Any email address that shows up in another field or in an error message is masked the same way.
  • IDs, scores, dates, counts, and the names of groups, courses, policies, and simulations are not masked, so the AI can still analyse your program and act on learners by ID.
  • To find one person, the admin names them in the chat. sibermate_list_learners with search compares that name or email with the real values on the SiberMate server and returns only the masked match and its ID.
  • Masking applies only to what MCP sends to an AI client. The REST API returns full records to your own systems.

Applying a spreadsheet without sending real emails

  • To change groups for many learners from a spreadsheet, add a column that masks each email with the formula below (Excel 365 and Google Sheets; email in column C), delete the real email and name columns, and give the AI only the masked emails and the target groups.
  • =LET(e,LOWER(TRIM(C2)),h,REDUCE(0,SEQUENCE(LEN(e)),LAMBDA(a,i,MOD(a*31+UNICODE(MID(e,i,1)),2147483647))),LEFT(e)&"***#"&LOWER(BASE(h,36,6))&MID(e,FIND("@",e),LEN(e)))
  • Names are optional, because matching uses only the email tag. To keep the sheet readable for people, replace each name with the masked form MCP uses, with the formula below (name in column A): Budi Santoso becomes B*** S***. Masked names carry no tag, so they never decide which learner a row matches.
  • =REGEXREPLACE(A2,"(\S)\S*","$1***")
  • If your spreadsheet uses semicolons between arguments (common outside English locales), replace the commas in the formulas with semicolons.
  • The AI passes the masked emails to the group tools, and the SiberMate MCP server matches each tag to a learner. If any tag matches no learner, or more than one, nothing is changed and the unmatched tags are listed, so a typo never moves the wrong person.

Employee data is entered in SiberMate, not in the chat

  • MCP has no tool to create learners or change a learner's name or email, so nobody needs to paste a roster or CSV with real emails or names into an AI chat.
  • Add people in the SiberMate platform (dashboard or CSV import), or sync them from your HR system with the REST API (POST /v1/learners, POST /v1/learners/bulk).
  • MCP then manages those learners by ID: groups, course and policy assignments, reminders, activation, and deactivation.
  • MCP also cannot create, change, or remove admin users, and cannot permanently delete learners, policies, or policy templates.

What SiberMate MCP stores

  • SiberMate MCP does not store tool results or learner data. Each call goes to the SiberMate API and the answer goes straight back to your client.
  • A usage log records each tool call: tool name, read or write, success or error, time, your tenant's ID, name, and domain, a hashed prefix of the API key, and a request ID. It does not record tool arguments or results.
  • Rejected requests (invalid credentials or rate limits) are logged with IP address and path, to detect abuse.
  • Error reports sent to monitoring have request bodies, cookies, IP addresses, and authorization headers removed.
  • OAuth tokens are encrypted and expire.

Read and write access

  • A direct API key and a standard OAuth connection get read and write access. Use a dedicated API key for MCP so you can revoke it without affecting other integrations.
  • Write tools that change assignments, send reminders, or remove data are flagged in the tool list, so AI clients can ask before running them.
  • To keep an AI client read-only, turn off SiberMate's write tools, or set them to ask for approval, in that client's connector or tool settings.

What the AI provider still sees

  • Tenant data that is not personal: group, course, and policy names, scores, risk and simulation results, and breached service names.
  • Anything typed into the chat, including the name or email an admin uses to look someone up, and any file an admin uploads there.
  • Masking works on known people fields and on anything shaped like an email. A person's name typed into free text, such as a group or policy name, is not detected, so keep personal data out of those names.
  • Pick an AI plan whose data terms fit your policy. Business and enterprise plans from the major providers generally do not train on customer data by default; confirm this in your provider's terms.

Last updated: 2026-10-19