Skip to content

StrataFS

A local, offline hybrid-RAG search server — BM25 + vector + metadata retrieval over local and cloud storage, exposed to AI agents as a native MCP search server.

Get Started View on GitHub

Install in 30 seconds

npm install -g stratafs
stratafs config init
stratafs serve

pip install stratafs
stratafs config init
stratafs serve

curl -fsSL https://raw.githubusercontent.com/neul-labs/stratafs/main/scripts/install.sh | bash
stratafs config init
stratafs serve

docker run -d \
  --name stratafs \
  -p 8080:8080 -p 8081:8081 \
  -v $(pwd)/data:/app/data \
  ghcr.io/neul-labs/stratafs:latest

Then search:

stratafs search "authentication middleware"
# or via REST
curl "http://localhost:8080/search?q=machine+learning"

Full-text (FTS5) and semantic (vector) ranking fused in a single SQL query. Natural-language queries return chunk-level results in under 100 ms on a laptop.

AI-native

A built-in Model Context Protocol server lets any MCP-aware agent query your filesystem as a structured resource — no glue code needed.

Multi-storage

Index local directories, S3, GCS, and Azure Blob with a unified, read-only interface. Each source is isolated in its own SQLite database.

Read-only by design

StrataFS never writes to your source files. All metadata lives in a parallel .stratafs/ directory. No permission escalation, no surprise mutations.

Compression-aware

Chunk content above 512 bytes is gzip-compressed at rest. Typical savings: 40–60 % disk overhead with transparent decompression at query time.

Cross-platform

Native binaries for macOS, Linux, and Windows. A Wails-based desktop UI ships alongside the daemon, with system tray integration on every platform.

Why StrataFS?

StrataFS sits between your storage backends and your AI consumers. It does not replace your filesystem — it augments it with a parallel semantic index:

  • No more grep -r — ask natural-language questions across your entire knowledge base.
  • Cross-file context — find related code, configs, and docs even when filenames don't match.
  • Zero lock-in — standard SQLite + FTS5, plain HTTP APIs, JSON-on-disk metadata. Walk away whenever you want.
  • Built for agents — the MCP server speaks the protocol your agent already knows.

Next steps

  • Quickstart — index your first source and run a search in under five minutes.
  • Concepts — sources, chunks, embeddings, hybrid search.
  • MCP integration — wire StrataFS into Claude, ChatGPT, or your own agent.
  • Architecture — a tour of how the pieces fit together.