mfs-server

The server side of MFS — Multi-source File-like Search: a context engine that turns code, docs, messages, databases, and object stores into one file-like, searchable namespace for AI agents and developers.

mfs-server is the heavy half of MFS. It owns the connectors, the ingest pipeline, the vector index, and the storage — and exposes everything over an HTTP /v1 control plane. The thin mfs CLI and the generated SDKs are just clients of this server.

• Project & docs: https://zilliztech.github.io/mfs/
• Source: https://github.com/zilliztech/mfs

Install

uv tool install mfs-server                     # core + local ONNX embeddings, Milvus Lite, SQLite
uv tool install "mfs-server[all-connectors]"   # add every connector's SDK
uv tool install "mfs-server[pg,s3,slack]"      # or just the connectors you need

uv tool install puts mfs-server on your PATH in an isolated environment. The defaults run fully offline: local ONNX (BGE) embeddings, Milvus Lite, and SQLite, with no cloud account or API key.

Run

mfs-server setup     # optional: write $MFS_HOME/server.toml (defaults to ~/.mfs)
mfs-server run       # all-in-one: API + inline task processing on 127.0.0.1:13619

For a horizontally scaled deployment, split the roles:

mfs-server api                    # HTTP control plane only
mfs-server worker --concurrency auto   # queue worker(s)

Point the mfs CLI (or an SDK) at the server and you're ready to mfs add a source and mfs search it. See the Quickstart for the full first run.

Optional native acceleration

A few hot paths — the gitignore directory walk, parallel content hashing, linear grep, and tail — have an optional PyO3 extension (mfs-server-rs). The server falls back to a pure-Python implementation when it isn't installed, so it is never required; it only makes large inputs faster. See Deployment for how to build it in.

License

Apache-2.0.