Skip to content

Status and limits

MFS is beta software under active development. The core loop — register a source, index it, search and browse it — is solid and used daily. This page is the honest summary of what's stable, what's still rough, and what isn't here yet, so you can decide what to rely on.

Pin what you test

During the beta, pin the CLI and mfs-server versions you're evaluating and keep them in step. Verify any API or SDK integration against the running server, not just the generated client docs.

What's stable

  • The CLI and the loop. mfs add, search, grep, ls, tree, cat, head, tail, export, and the connector / job subcommands are the steady surface. Examples in these docs track them.
  • Local first run. Install the CLI, uv tool install mfs-server, index a folder — the defaults (ONNX embeddings, Milvus Lite, SQLite, a generated token) work offline with no cloud account.
  • The /v1 control plane. protocol/openapi.yaml is the source of truth for endpoints and schemas, and both SDKs are generated from it. When auth is configured, every request except GET /healthz needs a bearer token.
  • Connectors. file is always available; the rest load when their dependencies are installed. Probe on the target server before a large sync.

What's still beta

  • SDK coverage. The generated Python and TypeScript clients cover the common server, ingest, retrieval, and browse calls. A few endpoints (connector management, the file manifest/upload steps, head/tail/export, job listing) are in the OpenAPI surface but not yet surfaced as generated methods — call /v1 directly for those.
  • API stability. The HTTP API may still shift before a stable release. Pin versions for scripts and integrations.

Scaled deployment

Running fully local is a first-class, supported path. The single-host container (Docker / Compose all-in-one) is also runnable today. A horizontally scaled split — separate API and worker processes against externalized Postgres, object storage, and a managed Milvus/Zilliz endpoint — is the documented direction the architecture is built for; treat the Helm chart as that target rather than a turnkey default. See Deployment.

Where MFS draws the line

A couple of things MFS deliberately is not, so you reach for the right tool:

  • It's not a mounted filesystem — no POSIX writes, locks, or kernel semantics. It adds search, browse, and read surfaces over sources.
  • It's not a vector database for your own app. It uses Milvus for its own index; talk to Milvus or Zilliz directly if you need a vector store.

See Why MFS for when it fits and when it doesn't.

What's ahead

Directional, not dated:

  • Multiple processing profiles — per-source pipelines, each its own collection, so a code source and a multilingual document source can use different embedding models.
  • Multi-user credentials and access control — so a hosted deployment can let each user bring their own sources and secrets.
  • A wider connector catalog.