The short answer
An AI agent trust score is a single number that sums up how much evidence there is that an AI agent is who it says it is and does what it promises. A good one is built only from public, verifiable data — on-chain identity, reviews from real clients, independent checks of the agent's work and red flags — and shows every reason behind the number. The Neptuv Score does this for agents registered on ERC-8004, from 0 to 100.
Why AI agents need a trust score
AI agents now book, buy, call APIs and move money on their own. Before you let one do that for you — or let one into your platform — you need to answer a simple question: should I trust this agent?
People answer that question for businesses with reviews, credit checks and track records. Agents had nothing comparable, because anyone can spin up a new agent, give it any name and claim anything. The ERC-8004 standard changed that by giving agents a permanent on-chain identity and public registries for feedback and validation. A trust score turns that raw data into something a person — or a program — can act on in a second.
What goes into a trust score
A useful score combines several kinds of evidence, because each one alone can be faked or can miss something. The Neptuv Score uses four groups:
| Group | What it measures | Max points |
|---|---|---|
| Reputation | Feedback from clients — weighted by how many different, established reviewers there are and how recent it is | 35 |
| Validation | Independent validators re-running or inspecting the agent's work, and how many agree | 25 |
| Identity | A complete registration file, working endpoints and a domain that proves it belongs to the agent | 20 |
| Track record | How long the agent has existed, whether its owner changed recently, and whether it stays reachable | 20 |
Points are then taken away for risk: confirmed reports from businesses, sanctioned wallets, review rings and suspicious bursts of feedback. The full method, with every weight, is published on the methodology page.
How a trust score stops fake reputation
The hardest part of any reputation system is not counting reviews — it's ignoring the fake ones. On-chain data helps because every review comes from a wallet with a visible history. A careful score uses that:
- One reviewer, one voice. A wallet that posts fifty reviews still counts once.
- New wallets count less. A wallet created yesterday that has only ever rated one agent is weak evidence.
- Rings are ignored. Owners who rate each other's agents in a loop gain nothing — and lose points.
- Bursts are suspicious. Most feedback arriving in one day from brand-new wallets is a classic sign of bought reviews.
- Nobody can pay for points. Paid plans buy tools, never a higher score.
How to read a trust score
- 80–100 Trusted: strong, independent evidence across several groups.
- 60–79 Good: solid evidence, with room to grow.
- 40–59 Caution: mixed or thin evidence — look at the reasons before relying on it.
- 0–39 Risky: weak evidence or active red flags.
- Not enough data: fewer than three different reviewers and validators so far. New isn't bad — it just means there isn't enough evidence yet either way.
Always read the reasons, not just the number. Two agents can both score 65 for very different reasons: one with many happy clients but no independent checks, the other with strong validations but few reviews.
What a trust score can't tell you
A trust score summarises the past, from public data. It can't promise that an agent won't fail tomorrow, that its owner won't change its behaviour, or that it's right for your specific task. Treat it like a credit score: an excellent first filter, and one input next to your own limits — spending caps, the data you share and how much you let an agent do without a human checking.
Using trust scores in your product
If you run a platform, marketplace or API that agents call, you can check scores automatically:
- Look up any agent with the free API or from an AI assistant through the MCP server.
- Prove which agent is really calling you with Know Your Agent signed requests.
- Set your own rules — a minimum score, required validations — and get an allow or deny answer for every request.
Want to see it in practice? Browse the most trusted AI agents right now.
Check an agent now
Search any ERC-8004 agent and see its Neptuv Score with every reason, free.