Every MindrianOS release in the order it shipped. Each one advanced a different axis of the system: memory, context, the graph, the Brain, the commands. The current release is v2.0.0-beta.1 “Build the Loop”, the one that taught Larry to say “I don't know”. The stream is the moat.
Solo founder with an idea, no team yet. Larry as a thinking partner, the Data Room as a second brain, 26 commands, knowledge graph dashboard, and PDF exports.
Team meetings happen, decisions scatter, assumptions go stale. v2.0 turns transcripts into structured intelligence with speaker ID, cross-meeting convergence, and contradiction detection.
Six phases in one day. MCP server routing all commands, Brain MCP hosted with API key auth, Opportunity Bank, Funding Room lifecycle, AI Personas, and embedded local graph. MindrianOS runs on every Claude surface.
The Brain is the moat. v4.0 locks it with API key management, then establishes the CLI visual grammar every interaction follows. Multi-room lets you work on multiple ventures. The autonomous engine lets Larry think independently.
The Data Room becomes visible. One command generates 6 self-contained views, a Canvas graph with particles, chat with Larry using your API key, one-click Vercel deploy.
Model routing per agent. Six new lifecycle hooks. Parallel agent execution. Spectral analysis revealing how Claude thinks. Design-by-Analogy via TRIZ + SAPPhIRE. Sentinel scans while you sleep.
The Room becomes a living intelligence hub. RoomHub detects project type and adapts. SnapshotHub exports 7 views as standalone HTML. 12-thread Constellation graph with De Stijl spectral coloring. Fabric Chat queries SQLite in natural language.
Causal reasoning with mechanisms, falsifiable predictions, natural experiments, DAG diagrams. The export system rebuilt for Synteris-quality single-file De Stijl hubs. 15 critical bugs fixed from the first external beta.
MindrianOS works identically across CLI, Desktop, and Cowork. The full command set served as MCP tools through 9 hierarchical routers. Brain-driven routing with 3-tier fallback. Token cost halved from 20K to 10K per turn. Skills compressed 65%.
Multi-room organization as a wicked problem. Centralized under ~/MindrianRooms/ with ICM identity files. Dual-graph layer (SQLite local + Neo4j Brain) holds ALL organizational axes. /mos:organize navigates the hierarchy with virtual projections and graph-informed grouping.
The Brain graph went from 4 framework chains to 125. From 4 stage mappings to 129. 444 semantic bridges connecting LazyGraph to canonical nodes. 7 leadership books codified. Every case study wired. Dummy-proof install with human-readable errors.
Based on SemNovel (Yale), the Brain embeds every room artifact and every framework into the same space, then finds empty regions. The interpretation layer, teaching intelligence applied to gap detection, is the potentially patentable moat. Discovery Cycle chains HSI, Whitespace, Reverse Salient, and Analogy.
MCP-native delivery. Brain client bundled as local stdio shim, two-step onboarding (install plugin, drop API key in ~/.mindrian.env). Auto-migrates existing HTTP transport. Co-designed statusline. Persona-aware first-touch surface. Dual-path detection routing uploads vs typed input.
Larry stops waiting to be asked. A 6-reach selector spine (context, contradiction, cross-room, brain, deep-research, hats) lets him reach for the right capability on his own and surface it as a Decision Gate. The new research engines (BONO six-hats debate, the unknown-unknowns blind-spot hunter, Trending-to-Absurd) run on one shared gated-chain executor that checks its own work before it files. Every decision becomes a typed edge in the local graph: the room is the receipt of the conversation. One front door (/mos:ignite) to start or excavate a room; /mos:discover runs Larry-led client, product, and JTBD discovery.
One coherent way to choose, everywhere. A single Shape F.0-F.9 selector vocabulary replaces ad-hoc prompts across every Decision Gate in the product, surfaced through a statusline co-designed around what your current job actually needs to see - room health, JTBD, context budget, quieter when you are reading, louder when something needs you. Eureka arrives as a genuinely new engine: rank cross-domain opportunity pairs across your whole room at portfolio scale, and when the local encoder is cold, get an honest lower-confidence result instead of a dead stop. Alongside it, a run of trust fixes: reach ranking that no longer disagrees with itself, room binding that survives concurrent sessions, a doctor that audits its own claims instead of asserting them.
The release that taught Larry to say 'I don't know'. The Brain becomes part of the install: one command mints an anonymous identity and connects a free 30-day trial, so method answers are grounded from the first session. Every answer that comes from the teaching graph names its source. And when the graph has no grounded answer, Larry refuses plainly and queues the gap for a future release, instead of improvising something that merely sounds like teaching. In beta, and said so everywhere.
Not the phase numbers. Not the architecture. Just what gets better for you and when. The Brain has its own roadmap. Larry gets sharper at every step.
The research engine that finds patterns across your rooms moves onto the same local embedding layer Eureka already uses, instead of a remote vector store. Faster, and one less external service between you and your own graph.
Turn a diarized pitch transcript into one Minto-structured feedback artifact, orchestrated across hundreds of submissions at a fixed per-unit cost ceiling. Local-only scoring throughout. Nothing reaches a real submission before a human confirms the judge's calibration - accuracy on a rubric like this is never self-certified.
$ npx @mindrian_os/cli