ideaverse-os
v0.1.1
Published
Bootstrap a position-addressed, LLM-agnostic knowledge vault from a template. Karpathy's wiki pattern with conversational ingestion baked in.
Maintainers
Readme
ideaverse-os
Your knowledge base, written by conversation.
Bootstrap a position-addressed, harness-agnostic knowledge vault in one command. Karpathy's wiki pattern with conversational ingestion baked in.
npx ideaverse-os init ~/my-vaultSite: ideaverse-os.ktncodes.com
What it does
| Step | Command | What you get |
|---|---|---|
| Init | npx ideaverse-os init <path> | A numbered folder skeleton (00-agentic-OS/, 10-cortex/, 20-/30- user domains, 40-raw/, 50-research-library/, 60-skills/, 70-daily/, 80-visualization/, 99-archive/), a 6-file auto_inject identity layer, and router files for every major harness (CLAUDE.md, GEMINI.md, AGENTS.md, .cursorrules, AIDER_CONVENTIONS.md). |
| Build | /ideaverse-os build (in your harness) | A 5-phase LLM-driven interview that compiles your compass, identity, workflow, domains, and optional layers into the lean files. Lock-file flow on every phase so you review before any mutation. |
| Capture | /ideaverse-capture <input> | Paste a YouTube URL, drop an Obsidian web clip, paste a plain URL, or write "I learned X today." Routes to the right handler, ingests, refreshes your cortex articles, and cross-links automatically. |
| Lint | /cortex-lint [topic] | Read-only audit of 10-cortex/ for stale verified_at, missing TL;DR, untyped sources, and triggered staleness signals. Diagnoses; doesn't mutate. |
Six bundled skills ship with init: ideaverse-os (the build interview), ideaverse-capture, cortex-compile, cortex-connect, cortex-lint, web-clip-report, yt-light-research. Each lives at 60-skills/<skill>/SKILL.md plus a copy at .claude/skills/<skill>/ for Claude Code's slash-command discovery.
Why
Existing knowledge systems force a tradeoff. ideaverse-os keeps what each got right and ships the missing piece:
| System | The tradeoff | |---|---| | Karpathy raw->wiki | No template. You build the structure manually. | | AI Impact infinite_brain | 16 node types, 10 edge types. Over-engineered for solo operators. | | Tiago Forte PARA | Designed for human re-reading, not agent retrieval. | | Cole Medin AI second brain | Static identity files. No ingestion model. Claude-only. | | Nick Milo Ideaverse-OS | Vault template only. No LLM build flow. | | ideaverse-os | Position-addressed scaffold + LLM-driven 5-phase interview + harness-agnostic ingestion. |
Install
Requires Node 20+ and a harness of your choice (Claude Code, Cursor, Aider, Codex, Gemini CLI).
# Bootstrap a new vault
npx ideaverse-os@latest init ~/my-vault
# Open the vault in your harness, then:
/ideaverse-os buildinit is idempotent and detect-and-prompt safe -- existing files trigger a per-conflict diff with overwrite / skip / abort. It never silently deletes.
How the LLM finds things
Every harness has its own convention file (CLAUDE.md, .cursorrules, AGENTS.md, ...). Init writes all of them, each pointing at the same lean 6 files in 00-agentic-OS/:
soul.md-- communication and behavioral rulesme.md-- your identity (role, tools, knowledge)user.md-- your workflow preferencescompass.md-- current priorities (refreshed weekly)memory.md-- session continuityrunbook.md-- the right command for the right intent
Every harness loads them at session start through its native auto-injection mechanism. One identity, every tool. See /ideaverse-os/concepts/harness-layer.
Cortex articles
Articles in 10-cortex/ follow a strict frontmatter standard so the LLM can rank, refresh, and surface decay:
---
title: "..."
description: "..."
topic: "..."
sources: [...]
created: YYYY-MM-DD
last_compiled: YYYY-MM-DD
verified_at: YYYY-MM-DD
confidence: low | medium | high
staleness_signal: "one-line condition that means this is stale"
---Sources use 5 typed-edge prefixes (supports::, contradicts::, extends::, mentions::, inspired-by::) so the relationship between source and claim is explicit. See /ideaverse-os/concepts/cortex-frontmatter.
Optional layers (Phase 5)
Three opt-ins, each independent. Skip any subset.
- Claude hooks -- session-start, load-memory, save-session shell scripts registered in
~/.claude/settings.json. Claude Code only. - GitHub backup -- private repo, baseline commit. Any harness; requires
ghauthenticated. - WSL paths -- fills
platforms.wslblock inpaths.json. Windows + WSL only.
See /ideaverse-os/concepts/optional-layers.
Templates
Phase 4 of build offers six starter shapes plus Custom. Each names your 20- and 30- domains.
| Template | For |
|---|---|
| work-personal | Default Western knowledge worker. Job + everything else. |
| creative | Artists, writers, makers. Practice + projects. |
| researcher-builder | Studying X while shipping Y. |
| trader-investor | Markets + life. |
| dayjob-sideproject | Income job + the thing you're building on the side. |
| solo | One operator, one venture. No domain split. |
Architecture rules
- The CLI is the source of truth. The Claude skill wrapper shells out to the CLI; the site documents what the CLI produces.
- LLM-agnostic core. Anything Claude-specific is Phase 5 opt-in.
- Detect-and-prompt safety. Init never silently overwrites. Every build phase writes a
.lock.mdfor review before promoting. - No heavy runtime. Node + Next.js + Tailwind. No Python for end users. No Electron, no Tauri.
Development
git clone https://github.com/ktnCodes/ideaverse-os.git
cd ideaverse-os
npm install
npm run build # tsc to dist/
npm link # then `npx --no -- ideaverse-os init <path>` works locallyContributing
See CONTRIBUTING.md. Issues and PRs welcome -- this is early.
License
MIT (c) 2026 Kevin Nguyen
