[Patina]
Patina
Proof of tenure for accounts a person already owns
Anyone can make a new account. Nobody can make an old one.
Live App •
Documentation •
MCP Server •
Verify a score
Overview •
The Score •
Privacy •
Architecture •
Quick Start •
Integration •
Test Results
Live Deployments
Overview
Patina reads the history in accounts somebody already owns and turns it into portable, verifiable evidence that a real human has been there for years. It is built on Vana, the data portability protocol: the data lives in the user's own Personal Server, is read once under a grant they can revoke, and never passes through an OAuth handoff that keeps it.
Sybil resistance is unsolved and expensive. An airdrop, a quadratic funding round, a DAO vote or a free trial all need to know whether they are talking to a thousand people or to one person with a thousand wallets. The usual answers are poor: biometrics are invasive, KYC destroys privacy, and most behavioural signals can simply be bought.
Somebody running an account farm can buy followers, bulk-upload posts and fill in a convincing profile. What they cannot buy is a decade.
Why Patina?
Patina answers tenure, not uniqueness. A freshly verified unique human can be produced today; sixteen years of corroborated activity cannot. The two are complementary, and the tool descriptions say so rather than implying otherwise.
Features
Ownership proved, not assumed
Collection is meant to run through Vana Desktop, which opens a browser on the user's own machine and asks them to sign in. Credentials never leave the device and Patina never sees a password.
But Vana has two collection paths and does not report which one it used. The other reads a public page from a profile URL, which proves an account exists and nothing about who holds it. So "desktop only" cannot be asked for; it has to be forced.
Every source therefore requires one scope a public page cannot serve, and the read is refused if it does not come back. LinkedIn must produce connection dates, Spotify saved tracks, YouTube its Watch Later list. A request containing something private cannot be answered from a public profile, so the person ends up on Desktop and has to log in. sources.test.ts fails if a source is ever added without one.
Steam was withdrawn for the same reason. Its connector never signs anybody in: it takes a Steam Web API key and a Steam ID, and a Steam ID is public. Anyone could have handed Patina a stranger's account age and friend dates as their own, and unlike YouTube there was no private scope to require instead.
Instagram is paused for a related reason. Its proof scope, instagram.posts, is one Vana's server collects from the public page, so a read proves the account exists and nothing about who holds it, and Instagram has no private scope Patina could require instead. It is not offered, both Vana routes refuse it before any money is spent, the handle resolver will not look it up, and Instagram data already stored is left out of every score. It comes back when the phone app's signed-in proof lands.
Eight sources, seventeen scopes
GitHub, LinkedIn, Spotify, YouTube, Amazon, Uber, DoorDash and Shop. All of a source's scopes go in one approval, so the whole manifest is eight approval trips rather than seventeen.
Timestamps in, content out
Every scope is requested for its dates. Captions, addresses, emails, track names, game titles and other people's names are discarded before anything is stored, and a test fails if any of them reach the store.
Signed, portable, and dated
Every non-provisional score carries an EIP-191 attestation anyone can verify offline against Patina's published address. No key, no OAuth, CORS open.
Attestations expire after thirty days, and the expiry sits inside the signed message rather than beside it, so checking freshness stays offline too. A revocation list would have fixed the same problem by making every verifier call Patina, which is exactly the dependency people integrate to avoid. Scores are signed with a key separate from the one holding the escrow balance, so a leak of one is not a leak of both.
Answerable about itself
/api/health does a real write-and-read against the database and a live read of the escrow balance, and returns 503 when the deployment cannot serve the next person, so ordinary uptime monitoring catches an outage without being taught what to look for. A password-gated /admin shows the funnel, the user numbers and how many more connections the balance will fund. Funnel counting is done server-side: no third-party script, no cookie banner, nothing that contradicts the privacy page.
Readable by agents
An MCP server exposes the score to AI agents mid-conversation, with a handle resolver that returns tenure without revealing identity.
The Score
Six components, weighted by how expensive each one is in time rather than in money.
Age and the two time components are earned outright. Depth, Vouches and Breadth are gated behind them: an attacker can manufacture volume, friends and breadth in an afternoon, so those only count to the extent that real elapsed history backs them up. A floor of 15% stops a genuinely young person being flattened to zero for the crime of being nineteen.
Scored against reference profiles, asserted in score.test.ts:
The gap between an ordinary person and a farm is 69 points. That gap is the product.
Honest limits
Stated on the site itself, not buried here:
- It cannot prove personhood. It reports how much real history it can see. That is evidence, not a certificate.
- A low score is not an accusation. Young, private or quiet accounts score low and are doing nothing wrong.
- Aged accounts can be bought. That turns a free attack into an expensive one whose price climbs with age. It does not make it impossible.
- Desktop proves session control, not ownership. A bought account arrives with its password. The bar moves from "knows a username" to "holds the credentials", which is a large jump and not the same word.
- The proof scope is a floor, not a guarantee. Requiring something private forces a real sign-in, and a real sign-in is still only a sign-in. What it rules out is somebody scoring a stranger's public profile, which is the failure that would make the whole number meaningless.
What is read, and what is kept
Every scope below is requested for its timestamps. This table is the trust argument, so it is on the privacy page in roughly this form too.
Timestamps collapse to month buckets before storage. The scorer only ever asks about months, so holding anything finer would be holding it for no reason.
Architecture
Connect flow
Read and settle
acknowledgeRead moves the request to completed, which is terminal and not read-ready. Firing it after the first scope would strand the rest behind a Personal Server no longer serving them, after the user had paid.
Scoring pipeline
Fragments are keyed by scope and merged at scoring time. A retried read overwrites its fragment rather than adding its months a second time, so a stuttering connection cannot inflate a score.
Quick Start
Prerequisites
- Node 20+
- A Vana app identity from account.vana.org/developers
- Upstash Redis credentials (optional locally; falls back to an in-memory store)
Installation
Fill in VANA_APP_PRIVATE_KEY and VANA_APP_URL, then register the app and fund its escrow:
Development
Integration
REST
Any public score, no key and no OAuth:
Verify an attestation offline
MCP
Test Results
The one worth knowing about: nothing sensitive survives normalisation feeds in a payload stuffed with a real-looking email address, two home addresses, a caption, a colleague's full name and LinkedIn URL, a game title, an employer, a song, a product and a restaurant, then asserts that none of the thirteen appear anywhere in what gets saved.
Project Structure
src/lib/score.ts is the whole argument. Every weight carries a comment explaining why it is what it is.
Tech Stack
License
MIT, see LICENSE.
Acknowledgments
Built on Vana and its Data Portability Protocol. Connectors come from PDP-Connect/data-connectors. Brand glyphs are from Simple Icons and Font Awesome's free brand set, both inlined so the app makes no external request.