open_scion

open_scion is an independent MIT implementation of the ideas in Anthropic's Redesigning source control post. The server stays online, serves sparse trees and single objects, answers blame and path history as memoized derived data, and lands changes with compare-and-swap instead of a long merge queue. State lives in SQLite or Postgres plus object storage; local Git repos are rebuildable caches. The server never calls a model — LLM conflict resolvers and reviews are optional and vendor-neutral.

Git hosting still assumes one human clones, works for hours, and pushes once. Agent fleets thrash branches with thousands of short edits, so full clones and merge queues become the bottleneck. open_scion is a public take on Anthropic's internal Scion design so teams can run agent-native source control without waiting for a closed product.

Clone the repo, make build, and run examples/graft/quickstart.sh for a temp server, cache daemon, two agents, and a land. Point graft CLI or the Go SDK at your config, use sparse clone for the paths each agent needs, and land through UserLand batching concurrent tip updates. Optional graft-verify bisects a broken tip and reverts the culprit; superfans watch the changeset stream; per-agent tokens and RPC allowlists gate what each agent may touch.

This is not another coding agent or Git UI. It is source control as an API for agent scale — sparse workspaces, CAS land, server-side history, and agent-first auth — while still speaking Git under the hood. Complementary to sandboxes like OpenShell or ax: those isolate runtime; open_scion redesigns how many agents share and land code on the same branches.

• API / SDK surface

• OpenShell ↗
• ax ↗
• AutoHarness ↗