LINZA

LINZA is a local MCP server for agent work with a Markdown folder and incoming text material. The server creates a SQLite sidecar beside the working folder: .linza/linza.db.

The sidecar stores the index, text chunks, imported artifacts, review proposals, accepted decisions, traces, and context packs. Source files remain the primary material.

Workflow

1. Connect a working folder through LINZA_VAULT.
2. Index Markdown files or ingest incoming artifacts.
3. Ask for a map, search, relation explanation, or review proposals.
4. Check evidence and dry-run output.
5. Apply only selected IDs.

Supported Now

• Indexing Markdown files from the working folder.
• Ingesting pasted text and local .md, .txt, .json, .docx, .xlsx, .pdf files.
• Semantic and lexical search over indexed notes.
• Review queues with stable IDs: rq-* for folder maps and aw-* for artifacts, memory, and calibr.
• Local memory, relation explanations, context packs, and trace review for agent work.

PDF extraction needs pypdf or PyPDF2. LINZA does not fetch web pages itself: an agent should extract readable text with a browser/web-fetch tool and pass it as an artifact.

Write Boundaries

Indexing, search, maps, relation explanations, and context packs do not edit source note bodies.

Visible changes go through review/apply. Apply actions return preview by default. A real write needs an exact ID and dry_run=false.

The normal reviewed apply path can write compact YAML for accepted role or domains. Hierarchy, causal links, memory, trace lessons, and incoming artifacts stay in the sidecar. Reports are written under .linza/reports; context packs are written under .linza/context-packs.

Next

• Use Cases - where LINZA helps.
• How It Works - what is stored and applied.
• Quick Start - installation and MCP setup.
• MCP Tools - public tools and agent_workspace.

Links

• GitHub
• PyPI
• Semiotronika Products