Storage that nobody owns, and nobody can read.
Slab encrypts your files on your machine, cuts them into erasure-coded shards and spreads them across independent operators. A node is one small process: no blockchain, no token, no wallet on the box. And the payment for your storage can't be linked to what you stored.
This page is stored on the Slab testnet and served from it.
What you can do today
Store, share, stream and publish.
Everything below runs on the testnet now, from the slab CLI or over plain HTTP.
Files and folders
Upload a file or a whole directory with slab put. Small files are packed together, so a folder of icons doesn't cost a slab each.
Share by link
A slab://root#key link is the whole capability. The key sits after the # and never goes over the network.
Stream and seek
Range reads fetch only the slabs a byte range covers, so video seeking works through any gateway.
Host a website
Point a signed feed at a directory and a gateway serves it as a site. This website is published exactly that way.
Mutable feeds
Owner-signed pointers that update in place: publish a new version and gateways following the feed serve it within 30 seconds.
Plain HTTP gateway
Any browser or curl reads Slab through /slab and /bytes. The paths follow Swarm's HTTP API (/bzz works too), so Swarm tooling points at Slab with a URL change.
How a file is stored
Four steps, and only the first one sees your data.
Encrypt
Your client encrypts the file with ChaCha20-Poly1305 before anything leaves your machine.
Slab
The ciphertext is cut into fixed-size slabs: 256 KiB today, 1 MiB for large files next.
Erasure-code
Reed-Solomon turns each slab into data and parity shards. Any k of them rebuild the slab.
Spread
Every shard goes to a different node, with caps per host (and next, per operator), so one failure costs one shard.
Durability
Why we're moving from RS 10+6 to 32+16.
10+6 survives any 6 of 16 shards failing. That sounds like plenty, but a large file has thousands of slabs, and only one has to be unlucky. Our model says a 1 GiB file on 10+6 is very likely lost when 10 % of nodes fail at once.
- 32+16 with 1 MiB slabs for files above about 1 MiB: 16 of 48 shards can fail, and it stores less (1.5× instead of 1.6×).
- 10+6 stays for small files and directory listings, where 1 MiB slabs would mostly be padding.
- Repair starts early: when fewer than 40 of 48 shards are healthy, seven more failures away from any loss.
Chance of losing a 1 GiB file
when a share of nodes fails at the same moment (exact model, slab-sim --scenario model)
Table view and whole-host failures
| Profile | 5 % nodes | 10 % nodes | 10 % of hosts |
|---|---|---|---|
| RS 10+6 | 2.4 % | 87.3 % | 83.3 % |
| RS 32+16 | 7×10⁻⁸ | 0.2 % | 0.09 % |
What's different
Things you don't get elsewhere.
Other decentralized storage networks are built around a token and a chain. Slab moves money off the data path entirely, which is what makes the rest possible.
Payment that can't be traced to your data
Vouchers are blind-signed (RFC 9474). The control plane sells a voucher but can't recognise it when a node accepts it, so no ledger links accounts to uploads. On chain-based networks, payments are public records tied to what they paid for.
A node that is just a storage node
No chain client, no token wallet, no RPC endpoint. It idles at about 25 MiB, roughly 25× less than a Bee full node, so it fits on a Raspberry Pi or a NAS next to everything else.
Euros, not a token
Customers buy vouchers at a price in euros; operators are paid for shards they prove they hold and serve. Nothing to speculate on, and no exchange between you and your storage bill. (Simulated on the testnet.)
Earn storage by storing
plannedOperators will be able to take earnings as storage credit: a node with spare terabytes pays for your own backups.
Nodes that update themselves, safely
Releases are signed feeds on Slab itself. Nodes stage an update outside their audit window and roll back if it keeps crashing. Operators don't babysit versions.
Storage made for AI agents
plannedAn MCP server with spending caps an agent can't raise, private feeds for agent memory, and forgetting by destroying a key.
Side by side with Swarm, Storj, Filecoin, Arweave and cloud storage →
Roadmap
Where it's going.
Operators first: durability, then nodes at home, then tools and a reason to run one. Agents after that.
Durability
32+16 for large files, operator failure domains, parallel slab reads.
Nodes at home
Hole punching and relays, so a node behind a home router needs no port forwarding.
Operator tools
A dashboard for your nodes, audits and earnings; bulk setup and staking.
Economics, then agents
A market price, storage credit, starting without a deposit. Then MCP and agent memory.
Have a disk you're not using?
The testnet is small and we're looking for the first independent operators. A node needs one UDP port, about 100 MiB of memory and whatever disk you can spare.