Open source · MIT · self-hosted
Git for MongoDB,
built for AI agents
Give each agent a branch of your business database. Review its changes, merge the work you accept, and undo mistakes. Experiments stay on their branches; applying a merge explicitly changes the target.
No login for the demo. Sample data, isolated to your session.
$ npm install -g argonctl@2.1.1local workflow · illustrated
$ argon sandbox create -p orders --name planner --ttl 1h
a local sandbox; run watch for CLI capture and schedule sweep for expiry
Published release: v2.1.1 · Branch databases, reviewable merges and agent workflows. Next: GCS chunk-store backend · Read-your-writes barrier for the wire proxy.
One order. Two agents. One decision.
Review changes to your business data
Conversation checkpoints do not capture changes to external orders and prices. This example makes the MongoDB data itself reviewable.
01 · Pin the input
Freeze orders at baseline: order-1 costs $49. Both agents fork that same pin, so their starting data is identical.
02 · Compare proposals
Planner changes the price to $44. Executor proposes $1 on its own branch. Both use ordinary MongoDB driver writes.
03 · Review before adopting
Inspect the planner diff and apply its merge plan. Main now contains $44. The executor plan surfaces a real conflict against that accepted change.
04 · Reject and recover
Undo the executor proposal on its branch, then discard it. You can also undo the accepted merge to restore $49; that recovery is recorded in history.
Run the complete Python example with the local quickstart. It asserts identical input, a real conflict, accepted data and the restored result. No LLM key is required.
Evidence, with its scope
0.86 ms
Historical metadata branch p50
479 B
Historical metadata bytes per branch
Diff → merge → undo
A review loop over your business data
The 0.86 ms / 479 B measurements use engine 8bf0f1e and a 50k-entry project. They measure metadata branching, excluding physical checkout, sandbox readiness and MongoDB storage. Read the versioned results and reproduction steps for current end-to-end measurements.
What to expect
Branch vs checkout
A branch is lightweight metadata. Checkout materializes a real MongoDB database; readiness, disk use and first-query costs depend on the dataset.
Isolation and merge
Experiments write to a separate branch database. Applying a reviewed merge explicitly changes the target branch. Use MongoDB credentials and network controls for access isolation.
Writer attribution
Native driver writes use the actor configured for the branch or run. Separate agents need separate branches; Argon does not identify individual clients sharing one connection.
Capture and undo
A healthy capture process records document changes. Undo requires complete images and retained history. Missing images or unsupported drop/rename operations mark history incomplete and prevent unsafe restoration.
Retention and pins
Time travel is available within retained history. Pins protect the data they reference while the pin exists; keep backups and choose retention for your workload.
Sandbox lifecycle
REST and MCP manage capture and TTL cleanup while the service runs. Standalone CLI sandbox creation requires a running watch process and scheduled sandbox sweep.
Anonymous demo
The hosted demo provides session-scoped sample data and review tools. Native MongoDB connections and physical sandboxes are available in your local deployment.
Try the complete review loop
Begin with the hosted sample, then run the same order workflow against your own local MongoDB replica set.