Bonsai is the architecture behind Sage. The idea fits in a sentence: ship a small base, and let it grow only the capabilities you ask for.
A bonsai gardener grafts a branch onto a rootstock and the tree makes it part of itself. Sage works the same way. The base image is the Rootstock™. Each capability is a Sprig™ it grafts on while running. The seam where they meet is the Graft Union™.
Most AI platforms ship one enormous image with everything baked in: every model and runtime anyone might want. You download gigabytes to use a fraction of it, and changing any part means rebuilding the whole thing.
Bonsai™ inverts that. The Rootstock™ owns what every install needs anyway: the web interface, accounts, storage, and the supervisor that does the grafting. Everything else waits in a catalog until you ask for it. A school that wants document search grafts three Sprigs™. A newsroom that also transcribes interviews grafts one more. Nobody carries what they did not choose.
The part we refuse to compromise on: Sage never fetches models or packages from third-party hubs behind your back. Your server pulls every Sprig from a registry we publish and runs two offline checks before opening anything. It verifies the publisher's Ed25519 signature, a check anyone can repeat with the stock minisign tool. It matches the package against a sha256 fingerprint pinned inside the base image you already run. Only then does Sage open and cache the Sprig. Grafts survive restarts and image upgrades. When a new Sage release pins a newer Sprig, Sage pulls it at the next boot. If that pull fails, the version you already have keeps serving.
One capability in one signed, verified package. A Sprig carries its own model weights or binaries, declares what it does and under what license, and speaks to the Rootstock™ across a plain process boundary. We publish fourteen today, from a 3 MB in-browser runtime to a 574 MB reranker.
Sage's Rootstock™ is licensed AGPL-3.0. Sprig™ authors ask what that means for them. Short version: your Sprig, your license.
No. A Sprig is a separate program that talks to the Rootstock over a published contract, across a real process boundary. That is the same arms-length arrangement as a proprietary app running on the Linux kernel, or a closed-source client talking to an open-source database server. Your Sprig can be MIT, Apache, GPL, proprietary, or anything else you choose.
Yes. The spec exists so you can build against a stable contract without touching Rootstock source code. A Sprig marketplace is on our roadmap, and our CTO has built one before: Alexander Somma was a founding engineer at Etsy (employee #5), the marketplace for independent makers.
No. Running an unmodified Rootstock triggers no AGPL obligation, no matter what you graft onto it. The obligation arrives only if you modify the Rootstock itself and expose the modified version to remote users, and it covers your Rootstock changes only. Your Sprigs stay yours.
Copying Rootstock source code into your Sprig crosses it. At that point your Sprig is a derivative work and AGPL applies. Build against the published contract instead and you never get near the line.
We are writing both sides of the Graft Union™ in the open, so a Sprig™ can come from anyone and a Rootstock™ does not have to be ours.
Both specs are v1 drafts under CC-BY-4.0. The GitHub repositories are the canonical text today. A spec hub is on the way at bonsai.sage.is.