Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/jsagir/mindrian-os-plugin/llms.txt

Use this file to discover all available pages before exploring further.

MindrianOS reads your workspace and writes only to your rooms (~/MindrianRooms/) and to session state (./.mindrian/). That is the entire write surface. Your notes, decisions, meeting transcripts, opportunity cards, and room entries never leave your machine. The Brain is a shared teaching graph that makes Larry sharper — it holds methodology intelligence, not your data. The boundary between the two is enforced in code, not just policy.

What stays on your machine

Everything you put into a room stays on your machine. MindrianOS does not transmit your venture data to any server, does not collect usage analytics, and does not have a telemetry pipeline for your content.
Data typeWhere it livesLeaves machine?
Room artifacts (decisions, opportunities, meetings, research)~/MindrianRooms/<room>/No
Local knowledge graph (nodes and typed edges)room/.room-graph/room.db (SQLite)No
Session state and binding./.mindrian/No
Intent cache and hook side-files./.mindrian/No
Persona overrides~/.mindrian/persona-override.jsonNo
Card-fire intercept diagnostic log~/.mindrian/card-fire-intercepts.logNo
Analytics / learning patternsroom/.learnings.mdNo
Framework names and problem-type enums in Brain queriesBrain API endpointYes — methodology only
Embedding model weights (one-time download by model ID)Local cacheModel ID only egresses

What goes to the Brain

The Brain is a Neo4j + Pinecone graph of methodology intelligence: frameworks, problem types, methodology chains, grading rubrics calibrated against hundreds of real ventures. When Larry queries the Brain, the query carries only generic methodology handles — framework names, problem-type enums, and Cypher patterns over node types that exist in the Brain’s own schema. Your room content never appears in a Brain query. A typical Brain query looks like this: a Cypher pattern over framework nodes (MATCH (f:Framework {name: 'Reverse Salient'}) RETURN f), or a Pinecone vector search over framework embedding space using a sanitized methodology term. Neither query type has a slot for user-specific content.
Canon Part 8 — the binding rule from MindrianOS’s own methodology constitution — states: “User data NEVER egresses to the Brain; it serves generic methodology only.” This rule is enforced by the Brain egress guard hook (see below), not left to convention.

The egress guard

The egress guard is a PreToolUse hook registered on every mcp__brain_.* tool call. Before any Brain MCP tool executes, part8-egress-guard-hook.cjs inspects the call parameters and checks for user-specific bytes — room text, artifact prose, personal identifiers, notes, or any other user-derived content. If user content is found in the payload, the call is blocked. The call never reaches the Brain. This is a constitutional enforcement: a user-data byte in a Brain call is not a warning, it is a blocked call. The hook cannot be disabled by environment variable or configuration. The inbound complement is brain-response-sanitize-hook.cjs, a PostToolUse hook on mcp__brain_.* responses. After any Brain call completes, this hook scrubs the response for any reflected content before it lands in the context window. The Brain might theoretically reflect back a query term; the sanitize hook ensures that even reflected content is reviewed before reaching Larry’s context.

The local knowledge graph

The local knowledge graph lives entirely in room/.room-graph/room.db — a SQLite database inside each room folder on your disk. This database stores all nodes (artifacts, claims, opportunities, entities), all typed edges (INFORMS, CONVERGES, CONTRADICTS, INVALIDATES, ENABLES, ROOT_CAUSES, DERIVED_FROM, and more), and all artifact metadata. The graph engine runs embedded — there is no graph server, no external database connection, and no sync process. A .gitignore pattern ships with MindrianOS that excludes room.db from commits by default. Your knowledge graph is yours, on your disk, and it does not go anywhere unless you choose to move it.

Brain query anatomy

When MindrianOS queries the Brain for methodology intelligence, the query is built from sanitized, generic handles only. Here is what a query looks like at the transport layer:
  • Framework lookup: a Cypher query carrying only the framework name (e.g., "Reverse Salient", "PEST Analysis") as a string literal matched against the Brain’s own node names. No room content, no artifact text, no user prose.
  • Semantic search: a Pinecone vector search built from a framework-name or problem-type-enum embedding. The embedding input is the sanitized methodology term, never a room artifact.
  • Chain resolution: a graph traversal query over FEEDS_INTO and PREREQUISITE edges between framework nodes. The query carries only node type and edge type filters.
The egress guard verifies this at call time. Any call that deviates from the sanitized pattern is blocked before it reaches the network.

Zero telemetry

MindrianOS does not phone home with usage data. There is no usage analytics server, no event stream, and no background reporting process. The only outbound network calls MindrianOS makes are:
  1. Brain API queries (methodology only, guarded by the egress hook)
  2. Web fetches for /mos:research <url> (the URL you explicitly provide, after you approve the ingest card)
  3. Academic and grants API calls for /mos:research and /mos:opportunities (the external domain allow-list in settings.json)
  4. The one-time model-weight download for the local embedding encoder (by model ID only, no room bytes)
Every other MindrianOS operation — graph derivation, opportunity qualification, grading, meeting intelligence, the knowledge graph, decision tracking — runs entirely on local bytes.

Local-first by default (MINDRIAN_MCP_FIRST)

The MINDRIAN_MCP_FIRST environment variable controls whether certain session-start and stop-gate operations are routed through the MCP daemon or run locally. When the flag is unset (the default), all logic runs locally — no daemon, no network hop, byte-identical behavior to every prior version. Setting the flag to 'cli' or 'all' enables the Phase 198 MCP adapter path, which communicates with the resident daemon over a local socket.
For maximum local operation, leave MINDRIAN_MCP_FIRST unset. The default keeps every hook and script running in-process on your machine with no daemon dependency.

Build docs developers (and LLMs) love