A Sprig is a capability your Sage grafts onto itself while it runs.
Document search, voice transcription, cloud backup, PDF export:
each one arrives as a single package from our public registry. Sage
verifies the publisher's signature and checks the package against a
fingerprint pinned inside the base image before opening it. Once
grafted, a Sprig survives restarts and upgrades.
Sage does not run pip install or quietly fetch models
from third-party hubs. Your server pulls from our
registry at
ghcr.io/sage-is once (or from
a mirror you run) and everything after that happens on your hardware.
Document search & memory
Everything Sage needs to read your documents and find them again later. Each piece arrives as a checked package from our registry.
The memory shelf. ChromaDB stores your documents as searchable vectors. Graft this one first; the embedding models below need it.
36 MB · Apache-2.0 · v2
Teaches Sage to read PDF, Word, and other document formats and split them into searchable pieces.
15 MB · MIT · v1
The lightweight default for English documents. Small and quick. The right starting point for most installs.
79 MB · Apache-2.0 · v1
Understands documents in about a hundred languages. Choose this when your knowledge base goes beyond English.
410 MB · MIT · v1
A stronger English-only cousin of MiniLM, for installs where search quality matters more than download size.
241 MB · MIT · v1
The same multilingual model as E5 large, served by one static binary. Choose it when you want the fewest moving parts.
561 MB · MIT · v1
A second pass that re-sorts search results so the best answer comes first. It speaks the same hundred languages as E5.
574 MB · Apache-2.0 · v1
Voice & media
Local speech-to-text and audio handling. Your recordings never leave your server.
Turns voice notes and audio files into text on your own server, in most languages. One static binary and one model file.
72 MB · MIT · v1
In-browser computing
Capabilities that run in your visitors' browsers, served from your Sage instead of a CDN.
Lets chats run Python safely inside the visitor's own browser. Your Sage serves the interpreter itself, so visitors never fetch code from a stranger's CDN.
48 MB · MPL-2.0 · v1
The WebAssembly engine that runs AI models inside the visitor's browser. It powers local text-to-speech and the evaluations leaderboard.
3 MB · MIT · v1
Export, backup & developer tools
The practical extras: turn chats into documents and back up your data. A third is for developers who want to hack on Sage itself.
Turns chats into shareable PDF documents, with fonts for Chinese, Japanese, and Korean included.
39 MB · LGPL-3.0 · v1
Copies your Sage's data to the cloud storage of your choice (rclone). Use it for off-site backups you control.
15 MB · MIT · v1
The full frontend build toolchain, grafted on demand. Lets developers modify Sage's interface without shipping build tools to everyone else.
297 MB · MIT · v2
How grafting works
You open Admin → Sprigs and click Graft.
Sage pulls the package, verifies the publisher's signature and the
pinned fingerprint, and roots the capability. Both checks happen on
your server, offline. Most Sprigs™ work right away without a
restart. When a new Sage release bumps a Sprig's version, Sage
pulls the new one at the next boot. If that pull fails, the version
you already have keeps serving.
This is why Sage's base image stays small: the capabilities live in
the registry and arrive only when you ask for them. The
architecture behind it is called Bonsai™, and
we are specifying both sides of it in the open: the
Sprig Spec™ for capability authors and the
Rootstock Spec™ for anyone building a host.
Ready to graft a Sprig? Open your admin panel and pick from the list
above.