Skip to content

Community projects

Open-source Scion clones only the slice you ask for

4 min read

Our AI analyst checked this story itself: it ran, measured and compared against the originals. What we checked

Written by an AI model from the linked sources and checked against them by a second model. How we work

All about: Coding agents →
Image: github.com

OpenScion, an open-source rebuild of Anthropic’s Scion posted to GitHub on October 13, clones a single directory by fetching just the trees on its path and the files beneath it. But the code so far runs only import, commit and clone, on a server that keeps everything in memory, so if you try it, restarting the server loses the repository. Server-side log, diff, rebase and landing are the roadmap’s next items, and persistent storage comes later.

Tiny Humans, which calls itself an AI lab, published it under GPL-3.0-only. The project follows the design Anthropic described in its post Redesigning source control, but it is not Anthropic’s code. Its 118 commits landed within 83 minutes that evening, and 111 of them list Medulla, Tiny Humans’ own tool for managing agents, as co-author.

Why agents break git’s model

Git assumes any client may go offline, so every clone carries the full history. That assumption fails in a monorepo where most clients are AI agents and CI jobs: thousands of short-lived, always-online jobs each need a small slice of a huge tree, and each pays for the whole repository anyway. At Anthropic, merges went from around 2,500 a day in May to over 30,000 a day by September. By then the main-branch history was almost 100 GB, CI ran about 25 million checkouts a day, and several hundred git mirrors were needed to keep up.

OpenScion’s README copies Anthropic’s answer. The server does the work: history, blame, commits, rebases and merges are meant to run there, while a client keeps a bounded cache and lays out files. A changeset lists only the files that changed relative to its first parent, so creating one costs the size of the change rather than the size of the repository. Blobs, trees and changesets are named by their BLAKE3 hash, so any cache tier can serve them without coordination and any client can verify what it got. Landing is planned as a server-side rebase plus a single-record swap of the ref, retried on contention, with no global merge lock.

Anthropic says Scion is by no means a suitable replacement for git for general use and only makes sense for enormous monorepos with swarms of agents.

What runs today

The client installs with cargo install openscion, which provides the opnscn binary, and cargo run -p openscion-server serves the API on http://127.0.0.1:7878. opnscn has five commands: version, server-info, import, commit and clone, where clone --path dir fetches one directory. The server answers seven HTTP routes, for storing blobs, fetching objects, creating and reading changesets, reading and moving refs, plus health and version checks.

Everything lives in one in-memory store. When we restarted the server after an import, main came back 404 and a fresh clone failed with ref "main" not found. The web app in apps/web is one page that shows the server version, protocol version and server URL. The ten most recent CI runs on GitHub all failed; the latest stops at Clippy, an error we reproduced locally, so the README’s gate of 90% line coverage in every source file did not run on that code.

Our analysis

What we checked

On October 18 we built OpenScion from source, ran its 180 tests, the quick start and the benchmark harness in a one-CPU container, and read the code, roadmap and commit history.

How we check stories

The README describes server-side history, blame, rebase and lock-free landing, plus filesystem and SQLite storage, as how the system works; the code has import, commit and clone over an in-memory store. That narrow part works as documented: all tests pass, and a cloned tree matches the original. The speed claim holds for import and clone, while at 10,000 files a one-file commit is slower than git.

FilesImportCommit one fileFull cloneOne-directory clone
1,000338.3 / 43.840.5 / 5.3880.0 / 97.5621.7 / 15.4
5,0002454.9 / 512.529.2 / 26.74476.7 / 238.9216.2 / 13.2
10,0002848.2 / 491.536.5 / 55.08265.0 / 456.7235.6 / 23.3

Times are in milliseconds, git first, on 4 KiB files over three runs. The authors measured on an Apple Silicon laptop, so only the shape compares; their table shows the same commit crossover at 10,000 files, 21.8 ms for git against 22.3 for OpenScion. Our one-directory clone of 100 files stayed between 13.2 and 23.3 ms as the repository grew tenfold.

The authors’ benchmark notes say the OpenScion side has no disk, no process starts and no packing, and that on a real network the full-clone gap would narrow or flip. Until 22:22 UTC that day, 35 minutes after the project was submitted to Hacker News, the README also said the comparison was not like for like and that OpenScion is not a general git replacement, the same limit Anthropic sets for Scion. A commit then replaced those lines with the sentence about importing and cloning several times faster.

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.