Jevbox
Self-hostable document library: Jev Choice beam search over folders, documents, and sections, plus Score usefulness filtering, with SpiceDB permissions and MCP.
Overview
Jevbox (github.com/extend-hq/jevbox, jevbox.extend.ai) is Extend's full stack document library from Andrew Luo, released October 2026. Better Auth organizations, Extend Parse indexing, SpiceDB authorization, and TypeSafe Jev hierarchical routing share one PostgreSQL estate. Search and automatic filing require a TypeSafe API key; README states there is no keyword or embedding fallback. Kushal Byatnal's launch post on X and the ModelSystem.One writeup describe Jev Choice beam search over folders, documents, and sections plus Score usefulness filtering before chat models answer with citations.
Problem: Team document libraries still lean on embeddings or keyword search that blur permissions, hide weak evidence, or need a separate vector stack to stay current.
Built for: Engineering and ops teams who want a self-hostable document drive with org sharing, cited chat, and agent MCP access without maintaining chunk indexes.
First indexed on Jev Directory: 2026-10-04
Creator and team
- Name
- Extend
- Handle
- @extend-hq
How Jev is used
- Role in the product flow
- Hierarchical Choice routing for filing and library search; independent Score usefulness on source passages before chat synthesis
- Primitives
- ChoiceScore
- State in
- Permission-filtered folder menus, document structural outlines, and overlapping source passages from parsed uploads (see docs/retrieval.md).
- Decision out
- Route log-probability scores across up to 4 beam paths; passage usefulness on a 4-level rubric with 1.5 floor and 2.75 early-stop threshold per retrieval doc.
- Parse uploads with Extend API or local text decoders; store sections and passages without generative summaries
- On upload, Jev classifies into existing folders or validates model-proposed branches (docs/organization.md)
- At query time, Jev Choice walks authorized hierarchy; weak evidence widens exploration up to documented caps
- Jev Score batches usefulness on candidate passages; chat providers answer only from accepted citations
Jevbox is the permission-aware document application lane, not DocJev's classify-and-split library and not jev-search's Search1API web planner. Extend stores structure and original passages, then spends Jev on routing and evidence quality instead of maintaining embedding indexes. SpiceDB checks run before provider calls and again on results, so revoked sources block dependent chats. MCP and REST expose the same search and answer tools with durable run handles documented in docs/api-access.md. Chat models from many vendors write prose; Jev never replaces them, it decides where to look and which passages merit context. Compare jkudish-jev-mcp when you only need judgment tools inside an existing agent, or docjev when your problem is boundary detection on one file.
Sourced performance claims
- docs/retrieval.md: beam retains 4 routes; Choice menus up to 64 children plus none; exploration capped at 96 node expansions and 96 passage scores per query.Source: github.com/extend-hq/jevbox docs/retrieval.md
- README: TypeSafe key required for hierarchical search and document questions; no keyword or embedding fallback when the connection is missing.Source: github.com/extend-hq/jevbox README Run locally
- Public GitHub repo extend-hq/jevbox had about 401 stars when this listing was drafted.Source: GitHub star count October 2026
- Extend's October 2026 launch post on X passed 900 likes overnight per ModelSystem.One news summary; treat engagement as social signal, not a benchmark.Source: modelsystem.one/news/extend-jevbox-document-drive
Features and stack
Features
- Extend UI Finder with grid, list, columns, and gallery views plus inline citation previews
- Automatic filing and upload-driven organization reviews over cached indexes
- Versioned REST API and MCP with API-key or OAuth auth under current document permissions
- Render blueprint, Docker Compose, Helm, and AWS/EKS Terraform deployment guides
- pg-boss workers for indexing, filing, chat, and MCP jobs separate from the web tier
Stack
- TypeScript
- Node 24+
- PostgreSQL
- SpiceDB
- Extend Parse
- TypeSafe System One
- Better Auth
Pricing: Self-hosted infrastructure plus your Extend parse, TypeSafe Jev, and chat provider keys. README documents Render paid services; no directory-hosted SaaS meter.
Links
FAQ
- Does Jevbox use vector search?
- docs/retrieval.md states the implementation uses TypeSafe as the decision provider and does not require PageIndex or a vector store. VectifyAI's jev-doc-search is the open recipe that does use PageIndex, for one long PDF. Jevbox is the shared library.
- How is this different from DocJev?
- DocJev classifies or splits single documents from natural-language rules. Jevbox is a multi-user library with sharing, chat, MCP, and hierarchical retrieval across many indexed files.
- Is the repository licensed for production forks?
- README encourages forks as a template and does not plan to accept pull requests. ModelSystem.One notes the public repo lacked a LICENSE file at launch; confirm license terms in the repository before you depend on it.
Related learn guides
Original Jev guidance that pairs with this product pattern.
- System One model
The model family behind Jev: parallel typed questions, one forward pass, probabilities you can threshold.
- Jev use cases
The patterns builders actually search for: moderation, routing, triage, RAG verify, and agent gates.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.