![]()
Security & approvals
Approving historical reads
Generic historical queries normally cover the last 90 days. Older or broader periods require your approval. Clients supporting MRTR and approval forms can show the server's elicitation request directly.
Chat approval is enabled on this server. The fallback below is available when your client cannot show approval forms.
When chat approval is enabled, other clients receive from query_entity
status: "approval_required", queryExecuted: false
and approvalRequired: true with an approvalRequest.
This is a request for your decision, not an empty search result: rows and paging
fields are absent. An executed query returns status: "completed".
The agent must show you the company, query scope, expiry and row limit, then
wait for your explicit yes or no in chat. No historical rows have been fetched yet.
After your answer, the agent calls confirm_read with the pending
requestId, companyId and approved decision.
If approved, it repeats the original query with the returned readApproval.
The same permission can cover subsequent pages; existing read limits still apply.
Chat confirmation trusts the agent to report your answer.
The server cannot independently verify that a person answered. This tool cannot
approve ERP writes. Do not put confirmed or requestState
inside query_entity arguments: those belong to the write preview flow below.
Native MRTR responses use requestState and inputResponses
on the tool-call params, outside arguments.
This server can post to the ledger
Two of the tools write accounting entries rather than master data:
create_quick_invoice and invoice_order. Both preview
first, and the preview is Uniconta's own computation - it asks
Uniconta to simulate the invoice and shows you the exact posting, VAT code and
totals it would book. Nothing is written until a person approves that.
An invoice number is drawn on approval and cannot be withdrawn; correcting a posted invoice means posting a credit note. An invoice is emailed to the customer only when their address is on the server's configured whitelist, which is EMPTY unless somebody deliberately filled it in. The preview says which of the two will happen, every time.
What your client has to support
This server supports MCP revision 2026-07-28 and reading with
2025-11-25. HTTP transport is stateless: no session id is issued.
Older clients can use initialize; modern clients use the newer discovery flow.
ERP writes require revision 2026-07-28 or later, MRTR (multi-round-trip requests), and declared form elicitation support with a human approval response. The protocol version alone is not sufficient. A client without these capabilities can read but not write. Broader historical reads can use chat confirmation only when enabled above. Chat approval never authorizes ERP writes.
Technical details: write approval exchange
The customer-write example below has two application steps. The first
previews and writes nothing; the second repeats the same arguments with
requestState and confirmed: true to request elicitation.
The client then completes the protocol exchange. Quick invoices have separate
approvals for draft creation and invoice posting.
1 tools/call create_customer {"account":"10001","name":"Acme A/S"}
headers Mcp-Method: tools/call
MCP-Protocol-Version: 2026-07-28
_meta io.modelcontextprotocol/protocolVersion: 2026-07-28
io.modelcontextprotocol/clientCapabilities: {"elicitation":{"form":{}}}
2 server -> {"inputRequired":true,
"requestState":"<opaque>",
"preview":"Will create customer account 10001 ..."}
3 server -> resultType: "input_required"
inputRequests: { "<id>": { method: "elicitation/create",
params: { message: "... Approve?" } } }
4 client -> inputResponses: { "<id>": { action: "accept" } }
5 tools/call create_customer SAME arguments,
plus {"requestState":"<opaque>","confirmed":true}