MNE-MCP: A Machine-Readable Public-Data Layer for Montenegro
MNE-MCP is a machine-readable national data layer for Montenegro. It connects 27 government sources and 1M+ raw records to AI systems through MCP tools, so answers can be grounded in legislation, companies, procurement, courts, finance, and public records.
The problem
Public information in a small country is not scarce. It is scattered. Laws, company filings, tenders, court records, and financial disclosures live across sites, formats, and update cycles. A person can find a document. A model cannot reliably find, cite, and re-find the same document unless the underlying layer is machine-readable and source-grounded.
Government AI that cannot point at a source is not infrastructure. It is a demonstration. MNE-MCP exists to make the first case possible.
What it is
MNE-MCP is not a chatbot. It is a data layer. Records are collected from 27 public sources, normalized into a form tools can query, and exposed through Model Context Protocol interfaces. The tools are built so a model can retrieve a passage and keep the source attached.
Domains currently in the layer:
- legislation
- companies
- procurement
- courts
- finance
- public records
Scale, as stated on this site’s résumé: 1M+ raw records. Related work includes Voice MNE-MCP, a voice interface for querying Montenegro’s legislation and public records naturally.
Architecture, in outline
Sources remain authoritative. The layer does not replace a ministry site. It makes the site queryable. Ingestion has to tolerate mixed formats, missing fields, and pages that were designed for browsers rather than for tools. Retrieval has to return enough context for a citation, not only a similarity score.
MCP is the contract with agents. A tool call is a permissioned, auditable action against a named source, not a free crawl. That is the difference between “the model looked something up” and “the system retrieved this record.”
What this note does not claim
This page does not invent source lists, latency numbers, or accuracy scores that are not already on the résumé. It does not treat a GitHub placeholder as a released repository. It does not assign a DOI that Zenodo has not issued.
Schema sample
A machine-readable description of the record types, without a dump of public records, is at /research/mne-mcp/record-types.json.
How to cite
Ivanović, Đorđe. 2026. “MNE-MCP: A Machine-Readable Public-Data Layer for Montenegro.” https://www.mrivanovich.com/research/mne-mcp
Citation metadata for software archives is in CITATION.cff in the site repository. After a Zenodo deposit, the DOI should replace the URL above.
Publication status
- Technical note: this page.
- Schema sample: published.
- Public source code: not released here.
- Zenodo DOI: not issued.
- ORCID link: not connected.