Control plane
Intent → any backend
Volumes, snapshots, clones, CephFS RWX and buckets over REST and gRPC, mapped to Ceph, NFS, ZFS and Longhorn by one StorageDriver trait. A fake driver runs the whole console with no cluster.
Architecture →4 + Nativestorage backends
150+REST endpoints
6DataBridge DB engines
2-wayDR with clean failback
Intent in. Backend details out.
Products request production storage, not pool names or CSI quirks. Atlas owns inventory, ownership, audit and the async job engine, and when you need a filesystem of your own, Atlas Native runs it on your nodes.
Control plane
One API, every backend.
Products state intent: volumes, snapshots, clones, CephFS RWX, S3 buckets. The StorageDriver trait maps it to Ceph, NFS, ZFS, Longhorn, macOS disks or Atlas Native. Real drivers never fabricate data, and every driver has a fake mode.
- REST, gRPC and SSE
- Quotas, audit export, OIDC/SSO
- DataBridge database migration
Atlas Native
Run the filesystem yourself.
Striped, checksummed extents, sharded Raft metadata, 3 replicas or Reed-Solomon k+m, cold extents tiered to S3, and NFS, SMB, S3 and CSI over the same files. One small pod for edge sites.
- Erasure coding with a rebuild controller
- POSIX ACLs and quotas
- Single-node edge profile
Agent-ready operations
Let an agent look. Not leap.
Atlas speaks the Model Context Protocol, so an AI agent can triage storage the way an operator does: the Ops Advisor, anomalies, eBPF I/O risk, DR status and jobs. It is read-only by default; the two opt-in writes (a snapshot, an alert acknowledgement) each need a single-use confirmation token, and MCP can never call out to an external LLM.
- 27 read-only gateway tools, 6 on the node agent
- Same roles, tenant scoping and audit as the REST API
- Opt-in guarded snapshot and alert ack; failover and delete never exposed
Disaster recovery
Fail over, fail back cleanly.
Two-way RBD mirroring and native replication. A non-forced promote is refused until this site has replayed the peer's demotion, so a failback cannot silently lose writes. Force stays available for real disasters.
- Live mirror status per site
- Promote guard
- Native replication with read-only replicas
- Automated application recovery drills
Try it
Would this promote be safe?
A model of the DR guard, not a live system. Toggle the state and try to promote site B.
POST /dr/mirrors/prod/promote
Refused: the local rbd-mirror has not replayed the peer's demotion yet.
Mode: journal. Mirroring status is read live from rbd mirror image status.
Atlas vs Rook alone
Keep Ceph. Add one API over all your storage.
| Atlas | Rook alone | |
|---|---|---|
| What it is | Control plane with drivers per backend, plus its own filesystem | Kubernetes operator that runs Ceph |
| Interface | Intent-based REST, gRPC and SSE | Custom resources and CSI PVCs |
| Backends | Ceph, NFS, ZFS, Longhorn, macOS disks, Atlas Native, S3 | Ceph |
| Cross-site DR | Two-way mirroring with a guarded failback; native replication | RBD mirroring CRDs |
| Database migration | DataBridge: discovery, full load, CDC, cutover | Not in scope |
| AI and agents | MCP server (27 read-only tools, opt-in guarded snapshot and alert ack), Ops Advisor, eBPF I/O risk | Not in scope |
Choose Rook alone when you only run Ceph inside Kubernetes and its own dashboard and CRDs are enough. Versus WEKA, including where WEKA is better today: the comparison.
Atlas Native
A filesystem of its own.
Striped, checksummed extents with sharded Raft metadata, erasure coding and rebuild, S3 tiering, and NFS, SMB, S3 and CSI over the same files. Lab-verified; NVMe and RDMA benchmarks pending.
Native roadmap status →DataBridge
Cloud to edge, on Ceph.
Six source engines, CDC, cutover, and Ceph-backed edge targets — migration as a control-plane product.
Quickstart →Agent-ready
Let an agent look. Not leap.
An MCP server with 27 read-only tools: inventory, the Ops Advisor, eBPF I/O risk, Ceph, DR and jobs, with the same roles and tenant limits as the API. Two opt-in writes, snapshots and alert acknowledgement, each behind a single-use confirmation token and an audit record. Failover and delete are never exposed.
MCP and agents →Day-2 and DR
Fail over. Fail back cleanly.
Alerts, maintenance, quotas, upgrade preflight, two-way RBD mirroring (snapshot, journal and pool modes) with a guarded failback, RBD consistency groups, and native replication to a second site.
See the console →
Open source. Apache License 2.0.
Atlas is Apache-2.0: run it, modify it, and ship it — in production, in managed services, and in your own products.
Read the licensing guide →
Running Atlas in production?
The software is free under Apache-2.0. Talk to Zyvor about deployment help, support, and integrations.
Book a demoPricing