About Metera

An agent reading an API has no way to tell a current number from a stale one, a measurement from a guess, or five providers from one provider wearing five names. It acts on all of them identically, because they arrive in identical shapes. The cost shows up later, in a decision nobody can reconstruct.

The industry answer has been to add sources, on the theory that more inputs means more confidence. That is only true when the inputs are independent, and independence is precisely the property nobody was measuring.

What we build

A judgment layer between an agent and the sources it depends on. It decides which source deserves each call, verifies the answer before it ships, and labels how far that verification got. It never claims a number is true. It reports what it knows, and it has not once overstated that.

The engine is a contextual bandit over decaying, staleness-inflated posteriors, with cheapest-first layered verification and failover inside the same request. In practice that means the routing improves continuously from real outcomes rather than from a static ranking, and a source that degrades is discovered by measurement rather than by a support ticket.

What we refuse to do

We do not take a provider's word for their own performance. Latency, uptime, freshness and agreement are measured by us and published as measured, which is why consent to be measured in public is a condition of joining rather than a checkbox in a settings page. The application form has no field for a provider to declare how good they are, deliberately.

We do not let anyone buy position. Routing is blind to price and to stake, and that is enforced in code and covered by tests rather than promised in a policy.

And we do not answer when we cannot verify. Refusing costs the caller nothing, which is what keeps refusing honestly compatible with running a business. A strict caller costs us more than a permissive one, and that is the correct direction for the incentive to point.

How it was tested

Against liars before the real world. Facing sources returning plausible wrong numbers that pass every structural check, wrong deliveries fell from 78.5% to 0.5%. Facing five providers sharing a hidden upstream, 575 confidently wrong consensus deliveries became zero. Two social providers agreeing byte for byte across eleven handles were correctly called one observation rather than two.

Then 500 live calls against independent exchange feeds, judged against a signed on-chain oracle: zero lies, zero unavailable, and a lying rate below 1% at 95% confidence, reported as an interval because a rate reported as a single number is not a measurement. Across every adversarial scenario the mislabelled rate was 0%.

What it runs on

Solana, with USDC settlement and x402 for payment. Anchors are Chainlink Data Feeds read from chain state, with Pyth behind them and the trace always naming which answered. Agents connect over MCP in one line, or over plain HTTP.

Where we are

Early, and specific about it. Several things in the navigation are marked as not yet built, and their pages say what exists today and what is missing rather than describing an ambition in the present tense. A company whose product is honest labelling does not get to be loose about labelling its own roadmap.