Parallel agents need lanes, not branch noise.
Independent AI work should be isolated, inspectable, resumable, and easy to stack.
Building in public
Sorrel is an open, agent-native version-control project for parallel development, portable workflows, and secrets-aware execution — without abandoning Git.
$ sorrel lane create agent/ui-copy
$ sorrel slice inspect app/landing
$ sorrel run local test-web
$ sorrel propose --with-secrets redacted
lane agent/ui-copy
changes: 7
touched: web/, docs/
runner: local-process
secrets: refs only
export: git-compatible
# proposal ready for review
Why Sorrel exists
Git remains the right compatibility layer for source history, remotes, and migration. But modern development now includes many parallel AI agents, cloud workspaces, generated code, selective context sharing, secrets, and portable execution.
Sorrel treats those concerns as first-class objects instead of forcing every agent handoff, review, runner, and environment decision through branches and ad hoc scripts.
Independent AI work should be isolated, inspectable, resumable, and easy to stack.
Agents often need a focused dependency closure, not the whole repository.
Local, container, and user-owned remote compute need a common portable job model.
Environment access, grants, and redaction belong inside the workflow layer.
Architecture
Proposals, reviews, projects, policies, runs, and team coordination.
Agent workspaces, dependency manifests, portable jobs, and secrets refs.
Content-addressed objects, blobs, trees, snapshots, changes, and future merge models.
Import, export, mirror, and migrate without locking teams into Sorrel.
Sorrel Core
Core is the protocol, object store, local runtime, CLI foundation, and SDK surface. It models snapshots and changes directly so agent-native concepts can be built without pretending every unit of work is just a Git branch.
Sorrel Hub
Hub is the server and UI layer for organizations, projects, proposals, review comments, workflow runs, policy, and agent coordination. It sits above the open core.
System primitives
Isolated streams of work for humans or agents, designed for stacks and resumable context.
Focused manifests that describe the files and dependencies an agent or review needs.
A JobBundle model for local process execution, containers, and future user-owned compute.
Secrets and environment management with refs, grants, local backends, and redaction.
Compatibility for import, export, mirroring, and migration so teams can adopt gradually.
Progress is tracked in the open as each module moves from skeleton to integrated system.
Migration path
Sorrel is not trying to make teams abandon Git on day one. The architecture keeps Git as an interoperability layer for source history, hosting, export, and escape hatches.
The goal is a better native model for agent work while preserving a practical path back to repositories, pull requests, and existing developer infrastructure.
Roadmap
Real object store, snapshots, changes, line-level diff, and history in the CLI — state persists on disk across runs.
Per-lane heads, lane switch, and an engine-level merge with merge bases, conflict markers, and stored Conflict objects.
Push/pull of content-addressed objects against sorrel-hub, with disk-backed object, ref, and metadata storage.
Collaboration flows referencing lanes and stacks, consuming Core policy; write flows in the Hub web interface.
Import, export, and colocated mirroring so teams can adopt Sorrel without leaving Git.
A stable library/IPC contract, TypeScript and Rust SDKs, then desktop and mobile clients on the same core.
Current status
The engine, the persistent CLI (including lanes and a real three-way merge), the policy spine with cross-repo conformance tests, workflows, vault, and Hub sync with disk-backed storage all work today. The docs pages go deeper into each layer, and the API reference is generated from the code itself.
Early interest