Verify before your agent signs.
One line to connect. Your agent hands Metera the exact transaction it is about to sign and gets allow or refuse, with the reasons and a receipt anyone can verify.
Metera is the verify + execution layer for AI agents, and its hosted MCP server is the shortest way in. Point a client at it and your agent gains nine tools: it can preflight a transaction before signing it, decide a swap against its own policy, read verified data, and watch an address for changes.
https://api.metera.xyz/mcp
Connecting
In an app, paste that URL as a custom connector. Metera sends you through a login and the connection is bound to your account, with no key to copy anywhere. In a terminal, add it as a server with a bearer token instead.
claude mcp add --transport http metera https://api.metera.xyz/mcp \ --header "Authorization: Bearer $METERA_KEY"
Transport is Streamable HTTP over a single POST endpoint. Both routes reach the same account and the same balance, so connecting an app by login and a terminal by header is one account, not two.
Preflight: the tool that refuses
preflight_transaction takes the exact transaction bytes your agent is about to sign, the intent behind them and your policy. Metera decodes the bytes, simulates them on two nodes and answers allow, allow_with_caveat or refuse, with every reason. Nothing is signed or sent. A refuse is free.
{
"decision": "refuse",
"decision_id": "9a14318e-48e0-4492-a423-5b25ee4fcffa",
"consensus_reasons": [
{ "layer": "layer1", "reason": "fee payer … does not exist on chain" },
{ "layer": "layer2_swap", "reason": "intent.pool … did not resolve to a real, recognized pool" }
],
"receipt": {
"receipt": { "verdict": "refuse", "message_hash": "sha256:…", "key_id": "ed25519:1886527ed07c735b", "…": "…" },
"alg": "Ed25519",
"signature": "…"
},
"debitedCreditsTotal": 0
}Every verdict, refuse included, comes with a receipt: Metera’s Ed25519 signature over what it decided, about which exact message. Anyone can check it with the public key at api.metera.xyz/v1/receipts/keys, without trusting Metera’s servers: how to verify one.
The other tools
get_verified_data fetches one datum from the best available source and returns the value, the level it earned, the source, and how fresh it is. It takes structured parameters rather than a sentence, because your model already knows what it wants and is better placed to fill them in than a second model guessing at the same question.
list_available_data returns the live catalog: every servable type, the parameters each one requires, and the highest level it can reach. It grows as sources are provisioned, so an agent can discover what to ask for rather than guess.
get_trace returns the full verification trace for an answer you were given: which sources were considered, what each scored, which attempts ran, and why one won. It is free to call, because it reports a call already made and already charged.
evaluate_gate composes a token risk check and a real simulated swap quote into an allow / allow_with_caveat / refuse decision on a Raydium CPMM or Meteora DAMM V2 trade — against a policy your agent supplies, not a threshold Metera decided for you. Read only because it only decides; nothing here signs or sends a transaction.
create_watch, list_watches, delete_watch and poll_watch tell your agent when a watched address changes. A change never is the answer by itself: it wakes a fresh verification, delivered by webhook or picked up with poll_watch.
None of the nine signs or sends a transaction, or moves anything you own. The watch tools create and remove subscriptions, and the manifest marks them as writes; the other five only read and decide. Execution is not available over MCP.
What comes back for data
Ask for a datum and the answer arrives judged. anchored means it was checked against on chain truth. consensus means independent sources agreed. single_source means it passed the checks but nothing confirmed it. Your agent reads the level and decides what to do with it, exactly as any other caller would.
{
"status": "delivered",
"value": { "token": "SOL", "priceUsd": 75.67 },
"level": "anchored",
"source": "okx",
"freshness": { "asOf": 1786582718269, "ageMs": 330 },
"traceId": "tr_9z23eb",
"cost": 0.0015
}When you ask for a level a type cannot reach, the call is refused before anything is spent, and the refusal names the highest level that type does reach. Refusing costs you nothing.