LibrisVault

Knowledge arrives as PDFs, office documents, web articles, images, and videos. Saving them somewhere is easy; what has value is the connected form - pages that link related ideas, cite their sources, and stay findable years later. Doing that by hand does not scale past the first busy week.

LibrisVault automates the whole path, across nearly any input channel: a watched folder, drag-and-drop, or anything sent to its Telegram bot from your phone - PDFs, office documents, blog posts and web links, YouTube videos, screenshots, plain notes. Everything is preprocessed and handed to sandboxed Claude Agent SDK sessions that write linked, cited wiki pages into an Obsidian vault. It all runs on your own machine - the vault stays a plain git repository on disk.

A connected vault raises a second problem, because now it knows what it does not know. Its pages end in open questions, point at pages nobody has written yet, and stop short where a stub stands in for an article. Chasing that by hand does not scale either. So LibrisVault also has the Fellows: standing research agents, semi-autonomous Claude Agent SDK instances that each hold a subject and a budget, work at night from what the vault itself says is unanswered, and leave their proposals and results for the morning, where you decide what runs and what does not. The intake fills the vault; the Fellows keep it moving.

The vault format and its ingest skill are claude-obsidian, an MIT-licensed project by AgriciDaniel, used unmodified. LibrisVault is the service layer around it - automated intake, sandboxed agent runs, a job queue, the dashboard - and it extends what the vault holds through documented extension points rather than by editing it: a registry that binds every page to one field of knowledge from a closed list, and a set of rules carried into every run, from English-only content to page hygiene and entity notability.

A Fastify backend over SQLite with a React dashboard, 134 source files on the server side and 126 on the web side. An ingest is a job in a queue. A job is an agent session. An agent session is a git commit. Nothing else in the design matters as much as that last equivalence: it is what makes every automated change revertible with a tool the user already trusts. A Fellow's night is a sequence of exactly those: each of its runs a session and a commit, revertible one by one.

A vault that stays a vault. Plain markdown in a git repository on disk, readable in Obsidian with the service switched off. There is no proprietary store and nothing to migrate out of.

A dashboard for the machine room. Intake, job status with live updates over server-sent events, a catalog of every page, and research over the web and the vault, each answer citing the pages it came from.

A shape you can see. The whole vault renders as one interactive wikilink graph, laid out by a web worker and drawn on canvas. Four overlays read it: Landmarks names a domain's most-linked pages and a reading order through them, Areas tints each detected community and captions it with the tags that set it apart, Spotlight isolates one community, and Bridges draws the links that cross between them, with their direction. The same layout colours by field of knowledge, page type, authority by inbound links or recency, and filters orphans and stubs. Sources are not a pile. Each one sits beside the concepts and entities it supports, so what a cluster of knowledge actually rests on is visible rather than asserted.

An order that survives growth. Every page carries exactly one field of knowledge from a closed, human-governed set. The catalog browses along that field, along page type, recency and backlink count. Orphans and stubs are filters rather than surprises. The gaps - links pointing at pages nobody has written yet - are listed rather than quietly dropped. What the validator finds stands as one counted row per defect, and every rule has a way out: a deterministic repair applied only to pages that still match what was approved, a bound agent run, or an accepted reason. A vault of a thousand pages stays navigable in the same shape as one of fifty.

Research assistants. A Fellow is an agent built on the Claude Agent SDK, with a subject and a configurable degree of autonomy. It proposes runs, each with its reasoning and an estimated cost, and works through them on a serial queue. The morning recap holds the results and the next steps it proposes, which you accept or veto. Its budget is a configurable share of your subscription, and every page a Fellow writes is an ordinary commit that git revert undoes like anything else. The Fellows are functional and still under close observation, so they ship switched off by default.

A library you can walk through. With the agents on, the vault renders as a room: one shelf per domain, filled to the page count it holds; one desk per Fellow, its screen lit while it works; a book cart where the maintenance runs work and the ingest queue's parcels wait; and a pinboard of every open question the vault's pages left behind, from which a question is handed to research or struck through on its own page.

One command to stand it all up. A setup script installs Node, the sandbox and preprocessing toolchain, the vault and the service, and leaves exactly one step for the browser: connecting an Anthropic account. Windows without WSL gets a PowerShell path that installs it first.

It is a service layer, not a vault format. The vault and its ingest skill are claude-obsidian's, used unmodified. What this project owns is everything that makes them run unattended: intake from three channels, the job queue, the sandboxed runner, the converter jail and the fence around what it produces, the domain governance, seventeen kinds of run, the research agents, the demo guard, the dashboard.

The sandbox is the boundary, not the callback. The Agent SDK's permission callback was measured to fire zero times, so the guarantee cannot sit there. It sits in the sandbox the runner configures: writes confined to the vault root, no web access in an ingest run, with the web reserved for the research flow and a Fellow's own runs; a Fellow's planning run is confined to what the vault already holds. If the sandbox cannot be created, the run fails rather than proceeding. Unsandboxed commands are switched off as well, because the shell tool ships a parameter that would otherwise let a command opt out. Without that one setting the sandbox would be decorative.

What is on this page is what ships today, the research agents and the library included since September 2026, behind one flag that is off by default. The next line of work is private for now.

The licence is deliberately not open source. PolyForm Noncommercial: free to self-host, commercial use by arrangement.