Power your agent
पब्लिशर का इंस्टॉल। PlugMCP एंडपॉइंट नहीं।
Host a copy with our keys and logs.
रजिस्ट्री में अभी कोई पैकेज्ड रिलीज़ नहीं है। रिपॉज़िटरी का README देखें.
रिमोट URL या आपकी मशीन पर कमांड? दोनों इंस्टॉल में क्या अंतर है
PlugMCP पर चलाएँ
A new server you create. It is not this listing's official endpoint. Use the publisher install when you want their URL and their credentials.
लॉन्चपब्लिशर से
Hybrid search over your own files — dense vector similarity, BM25 keyword matching, and metadata filters in a single query, running entirely on your machine. A Clojure/Lucene server holds the index; a Go CLI is the UX; the same binary speaks MCP, so Claude Code and friends can search your notes as a tool.
(אוֹצָר) is Hebrew for — and, fittingly, .
Local hybrid search over your own files — dense vectors, BM25, and metadata filters in one query. The same binary is an MCP server, so agents can search your notes.
otzar index -r ~/notes
otzar query "what did I decide about chunk overlap?"
# score path snippet
1 0.055 ~/notes/rag.md:42-47 Sentences are packed into windows of six, with two of…
2 0.031 ~/papers/rag.pdf p.4 Overlapping windows keep a claim and its evidence in…
Two files: the otzar client binary and otzar-server.jar — the whole server
packaged with its dependencies. Grab one of each from the
latest release; the jar is the same on every platform.
mkdir -p ~/.local/bin ~/.local/share/otzar
tar -xzf otzar_*_darwin_arm64.tar.gz -C ~/.local/bin
mv otzar-server.jar ~/.local/share/otzar/
You also need:
brew install openjdk or any Temurin
build will do.ollama pull bge-m3. It's the
default for its multilingual coverage — notably Hebrew, which most small
English-first embedding models handle poorly. Override with OTZAR_EMBED_MODEL.otzar finds both on its own. The jar: $OTZAR_SERVER_JAR, the directory holding
the binary, $XDG_DATA_HOME/otzar, ~/.local/share/otzar, then
<prefix>/share|lib/otzar for a binary in <prefix>/bin — so one directory with
both files works, and so does a /usr/local-style split. The JVM: $OTZAR_JAVA,
$JAVA_HOME, /usr/libexec/java_home, Homebrew's keg-only openjdk, then
PATH — probing each by running it, since macOS ships a /usr/bin/java stub
that exists and fails. otzar server status prints what it resolved, which is the
quickest way to check an install.
Building from source: see CONTRIBUTING.md.
The same binary is an MCP server, so an LLM host can search your index instead of guessing. To register it with Claude Code:
claude mcp add notes -s user -- otzar mcp # absolute path if otzar isn't on the host's PATH
It exposes three read-only tools — search, get_document, list_indexed_paths
— and starts the backing server itself, leaving it running when the host
disconnects. Writes stay on the CLI: an LLM should be able to find things, not
silently mutate the index.
Name it after what you indexed, not after otzar. The name becomes the tool
prefix the model sees: notes gives it mcp__notes__search, which reads like
something to reach for when you ask about your notes, where mcp__otzar__search
is a word that means nothing to a model. Pick papers, research, journal —
whatever your corpus actually is. Nothing depends on it; claude mcp remove <name>
and add it again to change your mind.
For the same reason, tell your host what's in there. The tool descriptions
describe the shape of a corpus because they ship to everyone; only you know it
holds ten years of org-mode journals. Two lines in CLAUDE.md close the gap:
My notes and reading live in an otzar index (~/Documents/Notes, org + markdown).
When I ask what I wrote, read, or think about something, search it before
answering rather than guessing.
Drop -s user to scope it to the current project, or use -s project to write a
.mcp.json that ships with the repo. claude mcp list shows what's registered
and whether it connects.
otzar index notes.md report.pdf # index individual files
otzar index -r ~/notes # descend into a directory
otzar index -f ~/notes # force re-embed even if unchanged
otzar index --now # re-scan everything already indexed
otzar query "vector database tradeoffs" # hybrid (default)
otzar query -m keyword "exact phrase" # BM25 only
otzar query -m vector "semantically similar" # kNN only
otzar query -k 5 -e md,txt -p ~/notes/ "topic" # top-5, filtered
otzar list # every indexed path, one per line (alias: ls)
otzar roots list # every tree you asked to keep indexed
otzar unindex ~/old-notes # the inverse of index: drop a tree from the index and
# forget its root (your files on disk are untouched)
otzar admin stats # documents and chunks in the index
otzar admin prune # drop indexed files that no longer exist
otzar admin clear # wipe the entire index
otzar admin clear --forget-roots # ...and stop tracking the trees it came from
otzar admin reindex # re-embed everything
otzar server status # health, index stats, which server it resolved
otzar server restart # picks up server changes
otzar server watch enable --interval 30m # auto-reindex on a timer (off by default)
otzar completion fish | source # also bash, zsh, powershell
A few things worth knowing:
index, query, list, admin *, otzar mcp — boots it if it isn't running,
with a one-line notice on stderr so piped output stays clean. The server
commands stay explicit, since a status that started what you asked it to report
on would be no use.otzar index -r ~/notes
is safe to re-run. otzar index --now re-scans every tree you've indexed —
picking up what's new, changed, or deleted — and server watch does the same
thing on an interval, without loading the model when nothing changed.otzar list is what is in the index; otzar roots list
is what you asked to keep indexed, which is what the watch sweep re-scans. A
clear empties the first and leaves the second, so the sweep would re-index
everything you just wiped — --forget-roots empties both.otzar admin prune.--mode picks the retriever; --ext, --path, and --after filter any mode.-j/--json on any command gives machine-readable output — for query, that
includes the hybrid score breakdown.Plain text, markdown, and source code (read as-is); HTML (tags stripped in
place, so line numbers still point at the file); PDF (via PDFBox — soft line
wraps rejoined into prose; no OCR, so a scanned PDF is skipped); and docx
(paragraphs from word/document.xml; the old binary .doc is not supported).
Anything else is ignored.
Each format declares what a hit can cite, because that differs — a line for text and HTML, a page for a PDF, a paragraph for a docx:
~/notes/rag.md:42-47 ~/papers/rag.pdf p.4 ~/notes/report.docx ¶12
Formats live in one registry, server/src/rag/index/extractor.clj, if you want to
add one.
Files are extracted, split into sentences with the JDK's BreakIterator (no
model file, and it handles Hebrew), packed into overlapping chunks, embedded via
Ollama, and written to a single Lucene index that holds the inverted index, the
doc values for filters, and the HNSW vector graph in the same segment. A query
runs BM25 and kNN with the same filters and fuses the two rankings with
Reciprocal Rank Fusion.
docs/design.md explains the choices and the alternatives that
were rejected; docs/api.md is the HTTP contract between the two
halves.
Working end to end and in daily use: hybrid retrieval, Ollama-backed embeddings, metadata filters, live indexing progress, the MCP adapter, auto-reindex, and index maintenance. Local-only, single-user, no auth — by design.
Test coverage: 150+ tests across the Clojure server and Go client, covering filesystem safety, extraction, indexing, retrieval, lifecycle, and MCP boundaries.
Personal project, best-effort maintenance: issues and ideas welcome, pull requests
probably not. See CONTRIBUTING.md to build and hack on it.
GPL-3.0 — see LICENSE.
Local hybrid search over your own files — dense vectors, BM25, and metadata filters in one query. The same binary is an MCP server, so agents can search your notes.
टूल
अभी टूल सूची नहीं है। रिमोट सर्वर को पहले हेल्थ चेक के बाद सूची मिलती है।
FAQ
इस पेज पर Cursor, Claude Desktop, Claude Code या VS Code के लिए stdio कॉन्फ़िग कॉपी करें। यह पब्लिशर का इंस्टॉल इस्तेमाल करता है, PlugMCP प्रॉक्सी नहीं।
इस रिकॉर्ड में कोई ज़रूरी एनवायरनमेंट वेरिएबल दर्ज नहीं है।
Otzar stdio पर लोकल चलता है।
पिछली जाँच 2026-10-28 को हुई। Local hybrid search over your own files — dense vectors, BM25, and metadata filters in one query. The same binary is an MCP server, so agents can search your notes.
