Windows · Local-first · Read-only MCP

Private knowledge.
Local retrieval.

Search Markdown, PDF, and Tika-supported documents with hybrid lexical and semantic retrieval—without uploading your knowledge base.

  • Indexing and queries stay on this device
  • CPU and NVIDIA CUDA supported
  • Open source under the MIT License
OmniClip RAG
Runtime ready
QUERY DESK What matters most in these notes?
Reranking · 4/6
Search your local knowledge base Search
Markdown PDF Tika Hybrid · CUDA
92
Decision principles

The best match is grounded in the original note and keeps its source path.

Markdown
Markdown · PDF · TikaOne search surface
FTS5 + LanceDBHybrid retrieval
BGE-M3Local embedding and reranking
Zero telemetryNo usage tracking

BUILT FOR PRIVATE KNOWLEDGE

A retrieval engine, not another chat shell.

OmniClip prepares precise, traceable context. You decide where that context goes next.

01

Hybrid retrieval

Exact FTS5 matches and LanceDB semantic candidates are fused, filtered, and optionally reranked.

02

Calm incremental watch

File changes are collected and updated in batches. When nothing changes, the watcher stays light.

03

You control the context

Inspect sources, select results, filter sensitive pages, and copy only the evidence you want.

04

Beyond Markdown

PDF and Apache Tika-backed formats use isolated indexes and background tasks that do not block the interface.

05

Resource-aware runtime

Auto mode can use NVIDIA CUDA, recover safely, and explain when a stage falls back or needs attention.

06

Read-only MCP

Let compatible AI clients search the same local index through two focused, read-only tools.

FROM FOLDER TO RESULT

Choose a folder. Let it index. Start asking.

Keep your folders and writing habits. OmniClip works around them.

  1. 1

    Choose your folders

    Add a Markdown vault and optional PDF or Tika directories. Each workspace stays isolated.

  2. 2

    Build the local index

    Create lexical and semantic indexes locally, with visible progress and resumable work.

  3. 3

    Retrieve grounded context

    Search in the desktop app, copy selected context, or expose results through read-only MCP.

CLEAR PRIVACY BOUNDARY

Local by default, explicit when data moves.

Indexing and search run on your machine. Model and Runtime downloads occur only when you start installation. Context leaves only when you copy it or an MCP client requests a result.

Stays local
Source files, indexes, embeddings, configuration, and query execution.
Moves only by action
Selected clipboard context or read-only MCP search results.
Network is optional
Used for user-initiated downloads of the app, Runtime, or models—not for telemetry.

READ-ONLY MCP BRIDGE

One local index, available to compatible AI clients.

OmniClip exposes status and search through a narrow stdio surface. It does not provide write, delete, or file-editing tools.

stdio · read-only
01omniclip.statusready · hybrid
02omniclip.search15 sourced results

Your client receives selected evidence. Your archive remains local.

DESKTOP INTERFACE

Every stage stays visible.

Follow a question from transparent retrieval stages to source-linked results you can inspect before use.

PRODUCT INTERFACE

Query Console and Results and Details

Ask in natural language, follow each retrieval stage, and review the evidence before preparing context.

Swipe the cards to see more interface states.

QUERY

Know what the search is doing

Semantic recall · 3/6
01
Decision journal / Opportunity cost

Matched by semantic meaning and source structure.

94
02
Notes / Trade-offs

Related context from a different source.

87
START

Runtime and index status

Runtime readyDetected
Semantic indexReady
Incremental watchWaiting for changes
EXTENSIONS

Isolated format pipelines

PDFBackground indexingReady
TikaOffice and document formatsReady

GET STARTED

Your knowledge base stays yours.

Download the Windows build, choose a folder, install the local Runtime, and build your first index.