Omnia Vault is a project brain you clone in five minutes. Your meetings, videos, code, and decisions become memory that Claude Code and Codex share, so every session starts warm.
It's the stack I clone at the start of every project in my own agency — 22 wired-together skills, 16 commands, and the hours and tokens they save me.
Works with Claude Code, Codex, or both. Python and git are the only requirements.
Every morning you re-paste the same context. Your agent re-reads the same files. The decision from three weeks ago is gone with the chat that made it. That's hours a week spent standing still.
I run an AI automation agency. Every client engagement starts the same way: I clone this vault, point it at their codebases and their recorded calls, and by the end of the afternoon my agents know the project better than a contractor would after a week of onboarding.
It is the single biggest time saver in my stack, and the biggest token saver. Agents stop re-reading repositories because they query a code graph instead. They stop re-deriving decisions because the decisions are written down with citations. The context I used to re-paste every morning is just there.
I built it for myself. It's the part of my operation I'd have been reluctant to hand over a year ago, which is exactly why it's worth putting a price on now.
One vault per client. Anyone on your team, or any AI, opens the folder, says "catch up," and is useful in two minutes.
Your codebase becomes a graph your agent queries instead of re-reading. Claude and Codex hand work off without you in the middle.
Your roadmap stops dying every time you watch a video about some new technique you should be using.
Record the calls, feed in the docs, and every recommendation traces back to the meeting it came from.
Meetings, docs, and videos compile into linked notes that cite their sources. "Why did we build it this way" takes seconds to answer.
Drop in a YouTube link. It checks the video against your roadmap and tells you what to adopt, what to skip, and when to just pass.
Claude plans, Codex executes, a baton file keeps them synced. On high-stakes work they argue first, and you only build what survives.
Code graphs, video transcription, design packs, report generation, and one command to install whatever's missing.
It opens in Obsidian, and the whole thing is plain Markdown you own. But a vault is storage. This is storage plus a compiler, a query engine, and a quality gate that refuses to let the notes drift from the truth.
CLAUDE.md and AGENTS.md mean Claude Code and Codex both arrive already knowing the rules.The workflows below exist because an agency breaks the things a solo side project never stresses: new people, moving requirements, and more codebases than anyone can hold in their head.
Drop in a standup recording. /video pairs a Whisper transcript with scene-aware keyframes, correlated frame by frame, then compiles it into linked notes. The full capture stays in Raw/Sources/ so every later claim can point back at the moment someone said it.
Each tracked repo gets its own committed graph in graphify/. Ask which modules a change touches, what the god nodes are, or how two subsystems connect, without an agent opening twenty files to find out.
The repo is the onboarding document. A new teammate clones it and reads the baton; a new agent session says "catch up" and reads the same one. Nobody waits on you to explain the project twice.
/roadmap locks vision, measurable success criteria, and phases with real exit criteria. Then /intel triages every new video or meeting against the active phase and returns a verdict: adopt, trial, watch, or skip. "Nothing here beats the plan" is a valid answer, and it stops the same idea being re-litigated next month.
Client work stays isolated. Imported repositories stay gitignored, chat archives never leave the machine, and the audit gate scans the working tree for secrets and machine-local paths before anything is committed.
/spar makes the two models argue over a plan in bounded read-only rounds before code exists, and whoever built a thing never grades it. /council convenes five blind seats for expensive-if-wrong calls, including one judging from an empty directory. Both keep the full argument on disk.
These are not twenty-two things you install separately and hope play nicely. They share one vault, one schema, and one quality gate, which is the part you cannot get by collecting skills yourself.
llm-wiki-ingestCompiles raw sources into short, linked notes that cite exactly where each claim came from.llm-wiki-queryAnswers from the compiled catalog before anything touches a raw transcript.llm-wiki-lintEnforces honesty: source links must resolve, counts must be accurate, tags must match folders.llm-wiki-maintainKeeps the wiki coherent as it grows instead of letting it rot into a note graveyard.project-context-queryThe three-layer answer engine: code graph, then wiki catalog, then one raw file.graphifyTurns any folder, repo, paper or video into a queryable knowledge graph, with query, explain, path and affected for impact analysis.war-roomThe living plan: phased roadmap with exit criteria, plan-aware triage of every new video, and a tool adoption ledger that tracks candidate through trialing to adopted, keeping the reasoning even for rejections.video-ingestRecordings and video URLs become correlated transcript and keyframe knowledge.import-projectAdopts an existing codebase and its loose files into the vault in one pass.defuddleStrips a web page down to clean markdown instead of saving the navigation and cookie banners.relayThe baton pass between Claude Code and Codex, so switching agents costs nothing.sparringAdversarial plan review in read-only rounds, then cross-graded builds.councilA five-seat decision panel split across both models, with anonymized opinions and blind cross-review.obsidian-markdownWrites the flavour of Markdown Obsidian actually renders properly.obsidian-basesBuilds database-style views over your notes.obsidian-cliDrives the vault from the command line.json-canvasCreates and edits visual canvases, mind maps and flowcharts as real files.tasteAnti-slop frontend design, so what you ship does not look machine-generated.ui-ux-pro-maxSearchable design intelligence: 79 styles, 192 palettes, 74 font pairings and 119 UX rules.humanizerStrips the tells out of AI prose before it reaches a client.stakeholder-update-writingTurns a week of commits into an update a non-technical client will actually read.github-pr-apiOpens and manages GitHub pull requests from machines with no gh CLI installed.How they connect. video-ingest writes captures into Raw/Sources/. llm-wiki-ingest compiles those into Wiki/ with citations pointing back at them. graphify indexes every tracked repo beside them. war-room triages new material against Plan/ and records the verdict with provenance. project-context-query answers across all three layers, cheapest first. relay carries the resulting state between Claude Code and Codex, and the lint and audit gate keeps the whole thing honest before each commit. That shared wiring is the product. The individual skills are just where it surfaces.
Two more stacks install through a guided command rather than shipping vendored: /design-setup adds the 23-command impeccable design plugin, and /docs-setup adds Anthropic's Word, PDF, PowerPoint and Excel document skills.
New project or existing one.
Meetings, videos, docs, your codebase.
Say "catch up" in either tool and pick up where you left off.
This came out of a real client engagement: recorded meetings, a multi-repo codebase, requirements that moved weekly. Before the template shipped, Codex reviewed it in two read-only rounds and found two real bugs. The review step earned its place.
MIT licensed. Your data stays on your machine. Delivered to your inbox right after checkout.