LINZA is a local-first MCP server. It is designed to help an agent read, index, and review a user-owned workspace without silently rewriting source notes.

• LINZA reads files only from the configured LINZA_VAULT directory.
• The sidecar database is created inside that vault at .linza/linza.db.
• Imported artifacts, chunks, approvals, calibr traces, and generated context packs stay local unless the user exports or shares them.
• LINZA does not include browser automation, cloud synchronization, telemetry, or a hosted memory service.

• Indexing, searching, artifact ingest, map generation, and calibr review are read-only with respect to existing note bodies.
• Apply operations are dry-run by default and use exact review item IDs.
• Generated reports are restricted to .linza/reports.
• Generated context packs are restricted to .linza/context-packs.
• write_file creates Markdown notes by default and blocks overwriting existing files unless explicitly allowed.

Treat every imported artifact as untrusted data. Web pages, chats, logs, documents, and PDFs may contain instructions intended to manipulate an agent. LINZA stores and analyzes these artifacts, but they do not become active rules, skills, memory, or note edits without review.

Security fixes are applied to the current public minor version. Before v1.0, that means the latest 0.x release line.

Please report security issues through the GitHub repository:

https://github.com/Semiotronika/LINZA-MCP/issues

Do not include private vault contents, API keys, or personal documents in public issues.

Expected response timeline:

• acknowledgement within 7 days;
• initial triage within 14 days;
• fix or mitigation plan as soon as practical for confirmed issues.