A drive for AI agents
A drive your agent can write to.
And read back tomorrow.
Storage that outlives the session, reachable by any model: one MCP connector, or plain HTTP for everything else. Files stay until you delete them, and each one has a URL you can hand to a person.
A real drive in one click. No signup form, email, password, or card.
The drive
Folders, previews, and one short name
A drive carries a short name like x7k2p9 and lives at a URL you can read out over the phone. Folders nest as deep as you want them to, and an agent files things straight into a path. Recipients preview images, video, audio and text without downloading a thing. Lock it and it asks for a passphrase.
Your drives
| Name | Size | Added |
|---|---|---|
releases | Aug 14 | |
press | Aug 09 | |
launch-teaser.mp4 | 48.2 MiB | Aug 15 |
press-kit-cover.png | 3.1 MiB | Aug 15 |
build-2026-08.tar.gz | 212 MiB | Aug 13 |
quarterly-report.pdf | 840 KiB | Aug 12 |
handover-notes.md | 12 KiB | Aug 11 |
press-kit-cover.png
- Size
- 3.1 MiB
- Added
- Aug 15
No invitations to accept, no permission matrix, no seats to buy. Open drives are public by name; locked drives require their passphrase.
Three ways this product gets used
If you are an agent
One line of config, then six hosted tools: drive, folder, upload, link, list, delete. The local binary adds account creation. What you write is still there on the next run.
If you are a person
One click opens a drive and creates the account for you. Folders, previews, and a short URL to hand out.
If you are receiving one
A page with the file's name and size. Preview it without downloading anything.
For agents · Model agnostic
One connector. Any model.
A drive belongs to your account, not to the assistant that happens to be talking. Add the connector in Claude, Cursor, VS Code, or anything else that speaks MCP, and the same files answer to all of them. Whatever cannot speak MCP can still speak HTTP: the same storage operations are documented as a public API, so no model is locked out.
create_accountlocal binary: a 16-digit number, created and remembered on first use.
create_drive: a short public name like x7k2p9, open or locked.
create_folder: a path like reports/2024/q3, every missing level created at once.
upload_file: path in, share URL out. Drops it straight into a folder if you name one.
share_link: a short-lived signed URL that downloads immediately.
list_files: drives, folders, files and sizes.
delete_file: moved to Trash for 30 days; its page closes now and issued links die at their short expiry.
Errors come back as {code, message, hint}: an agent that fails is told what to do next.
Run this
claude mcp add --transport http yogfile https://mcp.yogfile.com/mcp
Then say anything: on the first call it opens a page to connect your account (or create one). Nothing to install.
A session
> upload the build and send me a link uploaded app-v2.dmg 48.2 MiB drive x7k2p9/releases link yogfile.com/f/8c41… kept until you delete it
An example session, not a benchmark.
What agents do with it
- Keep state between runs
- a drive outlives the conversation, so the next run starts from what the last one left.
- Hand off between models
- one agent writes to a drive, another reads it. The name is the protocol, whoever is at the other end.
- Ship an artefact
- the build leaves as a link, not as a path on a machine nobody else can reach.
- Collect for a human
- several files, one short URL that previews everything.
The link
A stable page, a short-lived download
Every file gets a stable Yogfile share page. When someone downloads it, Yogfile mints a separate signed storage URL with a short lifetime and its limits inside the signature.
- Public by link
- anyone who has the stable Yogfile URL can open it while the file is active. It is not indexed, but the URL is access.
- Speed cap
- 10 MiB/s per connection, enforced by the storage node itself rather than by us asking nicely.
- Connection cap
- two per link, so one recipient cannot turn your file into a mirror.
Move the file to Trash and its stable page stops immediately. A direct storage URL already issued remains usable only until the expiry embedded in its signature.
Lifecycle
Files stay until you say otherwise
Nothing deletes itself here by default. When something should not linger, say so and the sweeper honours the date to the second. It is a policy you set, on the drive or on the upload, not a rule you live under.
- Nothing, by default
- a file you upload today is still there next year unless you delete it.
- A lifetime per drive
- set it once and everything dropped into that drive inherits it, including what an agent puts there.
- Or per file
- one upload, one date, for the thing that genuinely should not outlive the week.
The path of a byte
Four steps, and our API never touches your file
- 01
Ask
The client asks for an upload grant. The API checks the account and signs Ed25519 headers valid for minutes.
- 02
Send
The bytes go straight to the storage node the geo-DNS picked. They never pass through our API.
- 03
Confirm
The client reports the content hash. The file appears in your drive, ready to share.
- 04
Keep
It stays content-addressed and erasure-coded until you delete it or a lifetime you set runs out.
Underneath
The storage engine is ours, and you can read it
Yogfile is a thin product over Nauka, an open-source storage engine (AGPL) by the same author. The engine supports multi-node placement and geo-DNS. This launch deployment currently uses one physical storage server in Helsinki, so Yogfile is not yet geographically or hardware redundant. The interface shows the location handling your traffic.
- Reed-Solomon 4+2
- Nauka makes six shards and any four rebuild the file. Today they live on the same physical storage server, so this is not a backup.
- BLAKE3 content addressing
- a file's name is what is inside it, so identical uploads are one object.
- Authoritative geo-DNS, inside the binary
- the engine can answer DNS itself and route clients as nodes are added.
- Even our database sleeps in the cluster
- its snapshots ship into the storage it sells, so the service can be rebuilt from it.
Built by SYFRAH CONSULTING, engine first, product second. Nauka and the Yogfile MCP server are public source; the Yogfile API and web application are not. You can inspect the two pieces that store bytes and connect agents. Read the engine →
Pricing, honestly
Free while we launch
There is no checkout, so nobody is paying. High launch guardrails prevent accidental unbounded use; the speed and connection caps are the ones a normal user can realistically reach. If paid plans arrive, pricing and notice will appear before any charge. Here is every enforced limit today.
- Storage
- 1 PiB per account
- A launch guardrail, not a billable allowance. There is no checkout and no charge today.
- File size
- 1 TiB per file
- Bytes go straight to a storage node, so a big file is a long upload, not a refusal.
- File lifetime
- Yours to set
- No expiry unless you ask for one, on a drive or on a single upload.
- Drives
- 1,000,000
- A safety ceiling for the free launch plan, not a practical product constraint.
- Monthly egress
- 1 PiB
- Another launch guardrail. Downloads are not billed today.
- Speed per connection
- 10 MiB/s
- Enforced by the node itself, inside the signed link. This one you can actually hit.
- Connections per link
- 2
- So one recipient cannot quietly turn your file into a mirror.
- Signup form
- None
- One click still creates a 16-digit account. Optional MFA adds an authenticator and recovery codes. No email or card.
Hit a limit and the API tells you which one, and by how much.
Straight talk
What Yogfile is not
Not a backup. One physical storage server today. Erasure coding is not independent geographic or hardware redundancy. Keep another copy of anything irreplaceable.
Not a team workspace. No shared memberships or permission matrix yet. A drive belongs to one account and stays reachable through Finder, MCP and the HTTP API.
Not one assistant's feature. Any MCP client can add the connector, and anything with an HTTP client can skip it. Your files sit in your account, not inside a chat window.
Not recoverable. Lose the 16-digit account number and the account is gone. Recovery codes replace a lost second factor, not the account number. The number is a bearer credential, so MFA is the protection.
Not old. It launched this month. No logos, no uptime record, nothing true to show yet.
Not lawless. Reports are deduplicated. A threshold can restrict byte-identical content automatically, and that restriction can be reviewed, reversed, or confirmed.
Honest answers
Fair questions
Does this only work with Claude?
No. The connector is plain MCP, so any client that speaks it works: Claude, Cursor, VS Code, your own harness. The same core operations have a public HTTP API, so a model with no MCP support still reaches the same drive.
Do my files expire?
Not unless you ask. Put a lifetime on a drive or on a single upload and the sweeper honours it to the second. Leave it alone and the file stays until you delete it.
What if I lose my account number?
It is gone, and nobody can bring it back, us included: there is no email to reset with. Copy it somewhere real the first time you see it.
How do I protect the account?
MFA is optional and off by default, so the account number alone works until you enable it. Add a TOTP authenticator from the account menu and save the ten recovery codes. There is no email or support recovery if you lose both.
What is the catch?
It is free while we launch and there is no email address to sell. The only limits you can actually hit are the two on every link: 10 MiB/s per connection, two connections.
Can two different agents share a drive?
Yes, and that is the point. A drive is reached by its short name, so one run writes and another reads, whatever model is behind each of them.
Can anyone find my files?
There is no index, search, or directory, but public by link is still public: anyone who gets a file URL can open it, and anyone with an open drive name can browse it. Lock a drive to require a passphrase for its listing, and move a file to Trash to stop its share page.
Who is behind this?
One person, who also wrote the open-source engine underneath. That is the honest answer for a service this young.
What do you know about me?
A file's name, size, hash, and the lifetime you chose if you chose one. No email, no password. Account-creation rate limits can hold your IP in memory for up to one hour; abuse reports keep a keyed pseudonymous identifier derived from it.
Give your agent somewhere to keep things.
One line of config, no signup form, nothing to cancel. The first use creates a 16-digit account, and what the agent writes today is still there for the run after this one.