roadmap built, next, later, still reading
attest is the bottom of a stack: signed statements by people and the agents acting for them. Scoring sits above it, and above that, tools that tell you how likely something is to be real. Dated reasoning behind each item is in the repository's devlog; what actually shipped is on changes.
Built
- Passkey sign-up and sign-in; a per-site device key delegated by the passkey; revoke a session; second passkey.
- An AT Protocol account per person on our own data server,
did:plc, handle bound by DNS. - Records in the person's repository: vote, comment, statement, vouch, claim, verification, in our
monster.attest.*lexicons, each with the person's inline signature. - An index that follows the repository stream and verifies every record it counts.
- Profiles, vouching, proof of control over a site, domain, GitHub or Bluesky account, per-site pages, a drop-in widget, a JavaScript package, plain public reads.
- Tangled vouches, denouncements and follows, and Bluesky follows, imported as distinct trust edges for anyone who proves a Bluesky handle.
- Deleted and deactivated accounts drop out of counts. Lexicons published for resolution.
- Handle rules and reserved names; people's pages at
attest.monster/name; statements about a person recognised wherever made. - Bookmarks with tags and notes, a bookmarklet, a phone share target, tag pages.
- Sign in with an existing Bluesky or AT Protocol handle, asking only to publish attest's own record types; records go into your own repository.
- Agents: an agent asks with its own key, the person approves on attest with their passkey and may narrow it; a public permission in AT Protocol permission strings, with a daily limit and an expiry; records written via the agent are marked so and count as the person's; revoke any time. A command-line tool and a JavaScript client for agents.
- An admin page for admins, every request signed by their key; reversible hides of records and accounts; every action in a public moderation log.
- Offensive handles refused in twelve languages, with the word lists kept only as hashes (the handle-guard package), and names reserved for the site's own pages.
- A reading room built from bookmarks tagged
reading-room, a suggested shelf, and a history of the field. - Sign-in without a popup, for phones and home-screen apps: a redirect that returns the session in the address fragment.
- Nightly backups of every account, repository and the index, checked by restoring them.
- Where you sign in is private: sign-ins name the site only as a salted hash, published only once something signed with them is.
Principles the next work follows
- Trust is transitive and contextual. It flows from you through the people and communities you trust, weaker at each step, so second- and third-order connections count for less. It also depends on the subject: you may trust someone on climate and not on cooking, and tags give that context.
- Any node can be a community, and a person is one too. A human acts through a fleet of agents. Each agent has its own key and acts under a public permission the person's passkey signs, with limits and an expiry, revocable at any time. It earns no trust of its own; the person answers for it.
- Mutual beats one-sided. A vouch returned counts for more than a vouch given.
- No proof of personhood. No biometrics, no soul-bound identity. Whether an account is a person or an agent is not something attest tries to settle; "I have met this person" is an ordinary attestation, weighed like any other.
- Public by default, private where people need it. Everything attest publishes is public. Some relationships are rightly secret; how to hold those alongside is an open question below.
Next, in order
- Mutual vouches. Detect and show when a vouch is returned; weigh it above a one-sided vouch.
- Agents, the rest. A track record per agent, so a reader can see how a long-serving agent has behaved; revoking every agent at once; Claude's own account becoming the first agent acting for a person.
- Communities. A community account with its own page in the same namespace as people, founded and administered by its members. It issues memberships that each member acknowledges, so membership is mutual. Communities can vouch and be vouched for like anyone else.
- Contextual vouches. A vouch can carry tags, "I trust her on #soil", so trust is scoped to a subject from the start, as Pubky's tagged relationships do.
- Scores, first version. Transitive flow from a root you name, a person or a community, with a capacity limit at each step and decay with distance. Edge kinds weighted: mutual vouch, vouch, community membership, Tangled vouch, follows. Scoped by tag when asked. Every score shows the paths that produced it. Shown on pages, tags and the widget as "trusted by people you trust". The closest published design is Trustgraphs' vouch scoring algorithm: flow from chosen starting accounts, decay with distance, zero for anyone unreachable, integer arithmetic so every machine gets the same answer. We would rather adopt it than invent another, so attest's scores can be recomputed by anyone, and attest's vouches, which are AT Protocol records, could be an input to a Trustgraphs network.
- A person-held recovery key. Today the service holds the rotation key for accounts made here. Let people add a recovery key of their own that outranks ours, so their identity survives us. KERI's pre-rotation is the model to learn from.
- Rate our own capture risk in public, the way OpenHaven rates everyone else's: who holds which keys, which host, which company's protocol, and what we are doing about each.
- Import bookmarks from a Pinboard or Delicious export, or a browser's bookmarks file.
- Inbound trust from people not on attest, through a backlink index such as Constellation.
- Propose the URL-subject shape to the shared
app.certifiedvocabulary.
Later
A subjective score on any post, relative to you
Ask for any URL, any post, any record: how is this regarded by the people and communities I trust, and the ones they trust? The answer is a small scored graph with you, or a community you choose, at the centre, with the paths shown so you can see why. The same page viewed by two people gives two honest answers. The first version is in "Next" above; this is the fuller form, contextual and explorable.
Voting
Many votes do not need secrecy: a club choosing a date, a neighbourhood ranking street repairs, a community endorsing a proposal. Those can be done now with public signed records, one per key, with trust weighting to discount accounts nobody vouches for. We want to support several flavours: yes or no, ranked, approval, quadratic.
Secret ballots are a separate and much richer problem. A good digital election needs privacy, so nobody can see or buy your vote, and also retroactive proof, so you can later confirm your own vote was cast as you intended and counted. Those pull against each other, and the known solutions involve zero-knowledge and verifiable-shuffle cryptography we have not built. It is worth tackling eventually. Paper voting is slow and expensive enough that we vote rarely and on large bundles of questions; cheap, trustworthy digital voting would let people vote more often on smaller things, and would scale. That is a later chapter, and it will get its own research note before any code.
Place-based attestations
"If you are at this place at this time, something happens": a reward, a check-in, a proof of attendance, a witness to an event. These are usually built on blockchains. Public attestations can do much of it without a chain: a key reports where it is, and the claims are scored for plausibility. The cheapest defence against spoofing is physics. A key cannot be in Vancouver at noon and Lisbon at five past, so impossible jumps are discounted by a maximum travel speed, or by keeping a "ghost" that must walk the route. Witnesses nearby vouching for each other add more.
Claims like these invite adversarial checks, and that is healthy. If you say you were somewhere and were not, anyone with an interest, a competitor or a skeptic, can sign a statement that you were not, with whatever evidence they have. attest does not settle who is right; it records both statements, signed and public. Weighing them is a layer on top, and a hard one, because the challenger may be lying or mistaken too. Being caught in a false claim can cost standing in any score built later, which is the point of making claims in public. The service's job is only the attestations.
And the rest
- Authorship claims on content hashes, and a "check this file" page.
- Scores published as an AT Protocol labeler, so any subscribing client sees a root's view.
- Claims with resolution dates and a track record per key (Brier score).
- Institutions vouching for bylines, with the byline acknowledging it; pseudonymous authors vouched for by an institution.
- A browser extension that scores what you read, from your own root, without telling us what you read.
- An endpoint for AI agents that need to weight sources.
- Key export; an npm release of the client; the first sites to use it.
Still reading
Projects readers have pointed us at, or that we found, not yet studied closely enough to say how we relate. The reading room keeps them all in one place.
- OpenHaven: a human-verified catalogue of decentralized infrastructure, with capture-risk ratings and an open API. Its Identity & Trust section alone lists 57 efforts, among them FAN, GreenCheck, Pubky, KERI, did:webvh and Keyhive.
- The Decentralized Trust Graph work of Kaliya Young, Kevin Triplett and Drummond Reed, and the ToIP and DIF working group behind it: mutually signed relationship credentials, communities as first-class members of the trust graph, delegation to AI agents, and
domain/@nameagent names. Its diagnosis of the peer-only web of trust is ours too. Code: OpenVTC. - NorthStar Social (nstar.social): announced as a decentralized, local-first social network, with a beta announced for spring 2026; not re-read since.
- socialweb.computer: how its vouch and denounce policies are expressed, and whether it can read our records.
- The
app.certifiedlexicons: who stewards them, and how to propose an addition. - Community Notes's bridging method, for supports and challenges once volume exists.
Open questions:
- Private relationships. Some edges people will want to keep secret. Options: encrypted records only the two parties can read, relationships exchanged privately between wallets as in the decentralized-trust-graph work, or selective disclosure. Each costs something in a system whose scores are public.
- How a community governs itself: who admits members, how memberships end, and how graduated sanctions and conflict resolution (Ostrom's fifth and sixth principles) look in signed records.
- How far and how fast trust should decay per step, and how context from tags should combine with it.
- How strong should a follow be relative to a vouch, and how fast should an edge age?
- Who may resolve a claim, and how is a resolver trusted?
- Can reads stay private without a mirror on every device?
- Whether a passkey could ever control an identity directly. AT Protocol rotation keys need raw signatures, which passkeys do not produce today.
Know a project we should read? Tell us; the ways in say how.