An S3-compatible object store for small, self-hosted, geo-distributed deployments, designed so that a cluster can run outside datacentres on whatever second-hand machines are available. Each chunk of data is replicated in three zones; the cluster tolerates network failures, latency, disk failures and operator error; it ships as a single dependency-free binary and runs on modest hardware (any x86_64 CPU of the last ten years, ARMv7 or ARMv8, 1 GB of RAM, 16 GB of disk, 200 ms latency or less). It implements the Amazon S3 API so many applications already work with it, and it draws on research on Dynamo, conflict-free replicated data types and Maglev. It is run and funded as a community and public-funding project and is not a network anyone can join: the operator of a cluster decides who the nodes are .
LicenseCopyleft (AGPL-3.0). https://git.deuxfleurs.fr/Deuxfleurs/garage (Rust; the project's own Forgejo instance)
Dev Status 🟢 Active
OwnerDeuxfleurs (the project's repository and website are hosted on Deuxfleurs' own infrastructure; the project's funders are listed on its site)
Governing BodyDeuxfleurs, with its community; no formal governance document is described in the cited sources
Country Not stated by the project.
Start Year 2020
StackRust (single dependency-free binary); S3-compatible API; replication across zones; the site cites Dynamo, CRDT and Maglev research as inputs
FundingPublic funding: NGI POINTER (European Union Horizon 2021, three full-time staff for a year, 2021-2022), NLnet / NGI0 Entrust (one staff, 2023-2024) and NLnet / NGI0 Commons Fund (1.5 staff, 2025); donations and support contracts are invited
OriginsSoftware for small, community-run clusters: sysadmins who know the value of operator-friendly software, who do not have a dedicated backbone and want redundancy over the ordinary internet
DatabaseObject store with internal replicated metadata; no external database
Query LanguageS3 API (Amazon S3-compatible)
Data FormatsS3 objects and buckets; data stored in chunks replicated across zones
Mobile SupportNone: server software accessed through the S3 API
Web SupportCan host static websites from buckets, per the site's 'Host a Website' use case
Native AppsNone: server binary
TermsFree and open source (AGPL-3.0); support contracts are available on request
Funds Not stated by the project.
Based OnResearch on Dynamo (Amazon's highly available key-value store), conflict-free replicated data types and Maglev, per the site
P2P ArchitectureA cluster of nodes run by one operator (or a group of friends) across several zones or datacentres, with data replicated three times; not an open network
Overlay NetworkNone: the cluster connects over ordinary internet links between operator-chosen nodes
Content AddressingNo: objects are addressed by bucket and key through the S3 API
Local-FirstNo: it is a storage server for self-hosting rather than a local-first data layer
E2EENo: the server stores objects as given; encryption is the client's job
Byzantine Fault ToleranceNo: nodes are trusted members of one operator's cluster; the design targets crashes, partitions and disk failures
SignatureN/A: requests are authenticated with S3-style access keys
PermissionsS3-style access keys and bucket permissions managed by the operator
Semantic Web Compatibility N/A: no RDF, JSON-LD or SPARQL layer.
Smart ContractN/A: no ledger or contract component
Protocol Stack PositionApplication-layer (S3 over HTTP)
Asset / Value EmbeddingNone: no token or value-transfer mechanism
Protocol Maturity / StandardizationVersion 2.x production software that implements the de facto S3 API; no standards-body involvement
Update PolicyMutable: objects can be overwritten or deleted through the S3 API; versioning is not described in the cited sources
Storage Proof MechanismNone: the operator runs the nodes; replication in three zones is the integrity mechanism rather than audits or proofs
Economic / Incentive ModelNone: no payments or tokens between nodes; funded by public grants, donations and support contracts
Retrieval GuaranteeReplicated in three zones and designed to resist network failures, latency, disk failures and operator error; no external guarantee is offered
See something missing or that could be improved? Let us know →