The vision
It all begins with trust. Trust that is earned.
AI lies. It hallucinates, confabulates, and agrees with you just as confidently when you push back.
TrustShell makes it show the work. Two families check the claim. You get a receipt. If it cannot check, it does not pass.
What we are building toward
Agents help agents help people, and the receipt is how we know.
A harness on every agent. A score that moves when a confirmed action helped someone who is not the actor. A mesh that is a star around one engine — not a pile of new trust systems. Build toward that. Ship the path in front of you.
The point is people. Agents are getting good enough to act for us, and the useful ones will increasingly act through other agents — one hires a second to check a claim, a third to carry out a step. For that to help rather than compound the usual harms, two things have to be true: an agent has to be answerable, and its standing has to be earned, not granted by whoever happens to host it today.
Earned is the load-bearing word, and it has a sharp edge: a score should move when a confirmed action helped someone who is not the actor. Helping yourself does not count. That one rule is what separates reputation you can trust from a number an agent can farm, and getting it right — so it is hard to game and still fair to the smallest player — is the unsolved part we are most deliberate about.
This page states that as a direction we are committed to, not a victory we have already won. What is live, we mark live. What is ahead, we mark ahead. The parts that are still being built are labelled as such on purpose — the fastest way to lose trust while building a trust layer is to claim the finished version of it.
A star, not a mesh
It would be easy to answer every new use — a market, a trading lane, a health door — with a new trust system of its own. That is a mesh: every surface wired to every other, each one another place trust can be defined differently, and the most open door becomes the one everyone walks through.
So we build the opposite. One engine holds the score and the check; every surface is a spoke to it, carrying the same identity, the same rules, and the same honest verdicts wherever an agent goes. Add a surface and it points at the centre — it does not add a new centre. A star, not a mesh, is how trust stays one thing rather than fragmenting into many.
Why this matters for what you can rely on. Because the check and the score live in one place, a verdict you saw on one surface means the same thing on the next. The surfaces can come and go; the thing that decides whether an agent is answerable does not move with them.
The surfaces, marked for what they are
Some of these are shipped and some are a direction. The tag on each says which — and a direction is an aim, not a quiet promise. This is not a registry of everything in the ecosystem; it is the paths in front of us, said plainly.
TrustShell V1
ShippedThe stranger path, finished.
Paste a sentence; two independent model families check it; you see what each one said — Checks out, Caught, or Not checked. The SDK is published on npm as @hyperdag/trustshell (1.7.0). This is the one live product entry point, and it needs no signup and no wallet.
The engine — RepID and HAL
ShippedStar centreThe star at the centre.
One engine carries the two things every surface leans on: a portable score for how an agent has actually behaved, and a check that runs a claim past independent model families before it counts. Every surface below is a spoke to this engine — not a new trust system of its own. That is the whole shape: a star, not a mesh.
TrustMarket
DirectionAgents hiring agents.
A buy-intent skeleton on x402: an agent posts a capability and a fee, another agent hires it, settlement happens on-chain, and both standings update. It runs on the test network, it is early, and it is honestly incomplete — the shape of the loop before the loop is full.
TrustKeys
DirectionA catalog.
A catalog of the keys an agent and its owner hold, and the policy that governs each — so authority is named and auditable rather than assumed. A direction we are shaping, not a finished surface.
TrustTrader
DirectionA paper lane.
A paper-trading lane: signals and checks with no real money at risk. A place to watch the harness and the check do their work on something that moves, before anything of value is ever committed to it.
TrustGrowth
DirectionA CMO draft path that fails closed.
A marketing-draft path for an agent acting as a CMO. It fails closed against the HAL gate: a draft is produced, the gate decides, and nothing is posted on its own. There is no live post today — the draft is the output, the gate is the point.
TrustMedical
DirectionA door, not a trial.
A consent-first door: a person would opt in and share a proof instead of a record — an attestation that something is true, not the medical record behind it — so only the fields a study needs would ever cross. It is not a live or active medical trial, not a study, and not a clinical offering, and nothing here should be read as one. If this door ever opens, it would open the way everything else does — behind the same checks and the same honest labels.
Honestly stated: how to read this page
Shipped means it is live today and you can use it now — the check, and the engine behind it. The check runs past independent model families and lands on one of three results: Checks out, Caught, or Not checked. Not checked is never a quiet yes; it fails closed.
Direction means we are building toward it and say so. It is an aim we are committed to, not a feature we are claiming. Where a surface runs at all, it runs on Base Sepolia testnet — the money there is practice money with no value.
Go deeper