vaultcraft
A plugin for Claude, ChatGPT, Codex and Cursor that turns scattered materials (lecture slides, labs, meeting notes, research papers) into a linked, visual Obsidian vault you can study or work from the same day.
A real run in Claude Code: one command, the vaultcraft banner, a locked plan, then the build moves to a background agent.
At a glance
- Input: PDFs, slide decks (PPTX), Jupyter notebooks, Markdown and Word files, meeting notes, papers.
- Output: an Obsidian vault: atomic concept notes with a definition you can read on hover, lecture or topic notes with the source embedded, comparison tables, a start page and graph colours per folder.
- Vault types: studies, work, research and personal.
- Hosts: Claude Code, Cowork, claude.ai, ChatGPT, Codex and Cursor, from one repository.
- Price: free and open source (MIT). It runs on your own assistant account.
Why I built it
Notes pile up during a semester or a project, spread across PDFs, slides and documents. When you need them, before an exam or a client meeting, the information is there but you cannot find it fast. vaultcraft takes the material you already have and turns it into a knowledge graph built for recall: every concept one click (or one hover) away.
What you can do with it
- Build a study vault from a course. Slides, lab notebooks and readings become lecture notes, one note per concept and exam-style questions.
- Turn project docs into a work knowledge base. Meeting notes, decisions, people and companies, with a dashboard of open actions.
- Keep literature notes for research. Every paper linked to the concepts it introduces, every concept linked back to its papers.
- Add new material later. Point it at an existing vault and it only adds what is new; an interrupted run resumes where it stopped.
- Choose how it writes. Detailed narrative, study sheet, a fixed template or reference style; three depth levels; English, Polish, German or Spanish; explanation styles such as ELI5, worked examples, history and real-world cases.
Better context for AI, not just better notes
A vault is also the best input you can give a model. Clean Markdown with one concept per note, front-matter metadata, definitions up front and explicit links is the format language models read most reliably. Point Claude, ChatGPT or any agent at the vault instead of the raw PDFs and slides, and it finds the right note through the links, reads a few short files instead of whole decks, and spends far fewer tokens getting to a better answer. The same structure that makes revision fast makes retrieval cheap and accurate.
What you get
Every lecture note opens with its metadata and the original slides embedded inline, so the summary sits right above the source.
Comparisons that come up in exams get their own table, with a one-line answer you can say out loud.
Concepts link to each other and to the lectures they come from, so the graph shows how a course fits together. This one is a four-course vault.
How a run looks
The agent gets every intake answer as a locked parameter, so it never asks twice, then works through its phases.
- Intake. A few questions: vault type, format, depth, language, explanation styles, sources and where to save. Answer "defaults" to skip them.
- Goal. Before writing anything, it records a measurable goal: which sources to cover, which notes to produce, and the check that says the vault is done.
- Extract. It reads every source, saves the figures, and lists every named technique, model, metric and formula in it.
- Write. Concept notes, lecture or topic notes, comparison tables and a start page, linked both ways, with the source documents embedded.
- Check and fix, until done. A check script measures the goal; whatever it finds goes back to the writers, and the run does not finish until the check passes.
A team of agents, each on the right model
In Claude Code and Cowork the work is split across agents, each pinned to its model, so the expensive model only makes decisions and the cheap one does the volume.
| Agent | Model | Job |
|---|---|---|
vaultcraft | Your session's model | Plans, builds the concept inventory, sorts the work, writes briefs, judges the check |
vaultcraft-drafter | Sonnet | Exemplar notes, hard concepts, lecture walkthroughs, comparison tables |
vaultcraft-extractor | Haiku | Reads each source: text, slides, figures, named items |
vaultcraft-writer | Haiku | Writes the bulk of the notes from briefs and applies fixes |
It works in a loop, one batch of sources at a time: extract, decide, write, check, fix, record progress. An item the Haiku writer cannot fix goes to the Sonnet drafter, then to the orchestrator. Progress is saved in the vault, so an interrupted run continues where it stopped.
A real test run
Three NLP lectures from the repository's demo materials, built in Claude Code with every default:
- 67 notes: 3 lecture walkthroughs, concept notes for every named technique, comparison tables and a start page.
- The goal check passed: 0 uncovered items, 0 broken links, 0 orphan notes, every concept with a definition and an example.
- Cost on API prices: $3.44 in total. The Haiku writers produced most of the text for $0.14.
The goal check
check_vault.py ships inside the plugin and needs only Python's standard library, so it runs on your machine and in the ChatGPT and claude.ai sandboxes at no extra cost. It reads the vault the way Obsidian resolves links (case-insensitive, aliases, embeds) and reports:
- named items from the sources that have no note yet;
- broken links and notes nothing links to;
- concept notes without a definition, with a definition too long for the hover preview, or without an example;
- lectures that do not list the concepts that name them as their source.
Run it on any vault: python3 check_vault.py <vault>.
Where it runs
- Claude Code and Cowork: the skill runs the intake, then the agent team builds the vault straight into your folder.
- Codex and Cursor: the skill runs the whole pipeline and writes to your folder.
- claude.ai and ChatGPT: it builds the vault in the chat sandbox and hands you a zip to open in Obsidian.
# Claude Code
/plugin marketplace add MikolajSapek/vaultcraft
/plugin install vaultcraft@vaultcraft
# Codex
codex plugin marketplace add MikolajSapek/vaultcraft
codex plugin add vaultcraft@vaultcraftListings in the Claude, ChatGPT and Codex, and Cursor plugin directories are in review. On claude.ai, the zip from the latest release can be added under Customize, Plugins, Upload plugin.
The optional server
The ChatGPT and Codex directory currently lists only plugins that include an MCP server, so that version adds a small one at vaultcraft.sapcorp.eu with three read-only tools: Obsidian settings for a new vault, the note templates, and the same link check as the script. It stores nothing and logs only failed requests. The Claude and Cursor versions never contact it, and the skill does the same work without it.
How it compares with Smart Connections, Copilot for Obsidian and Notion AI: AI note taking tool for students: how to take notes with AI.
Data and privacy
- It reads only the files you point it at and writes only inside the vault folder you choose.
- No account, no telemetry, nothing runs on install. It asks before installing any tool.
- Web searches, made with your assistant's own search, fill gaps in a concept and are cited in the note. Your files are not sent.
- Full details: privacy policy and terms of use.
How it is built
- One source of truth for the pipeline, shared by the skill and a generated agent, with tests in CI that fail when they drift apart.
- Each agent's model is fixed in its own file and checked by a test, so the cost routing cannot silently break.
- The goal check and the server's link check are two implementations of the same rules; a parity test runs both on the same vaults and fails on any difference.
- One repository packaged for four hosts, each with its own manifest and the limits its directory enforces.