@solsniper330@useautonomo For Base USDC, can the agent retry the original request and get its result back if the payment settled but the response was lost?
I’d avoid signing a second payment until the first one is reconciled.
We built provenance into Testril so you can check our numbers against the chain.
For this demo, we used it to rebuild the transfer list directly from Ethereum blocks.
The script is public if you want to check our work.
We were building a demo comparing Nansen and Testril pricing and hit a snag: same question, very different answers. 7,139 USDC transfers vs 292 for the same hour on the Uniswap V3 USDC/WETH pool.
Turns out Nansen's summary was covering the whole day, not the hour we asked for. Nothing in the response said so.
So we checked Testril's numbers the hard way. Using Testril's provenance report, we rebuilt all 795 transfers straight from Ethereum blocks with our own script. Every one matched. The code is public if you want to rerun it.
If an agent is paying for data, it needs a way to check where the answer came from.
https://t.co/opT56QBHbM
Try this with your agent:
“What was this wallet’s USDC balance at block N on Base? Show me the source for that block.”
Pick a wallet and block, then use the Testril recipe to check coverage and get a quote before reading.
https://t.co/nxZaomxmuc
@BraveSearchAPI This makes sense for an agent that only needs a handful of searches.
We’re doing something similar for onchain data. If another agent has already indexed the blocks you need, you can just pay to query it on demand.
@xona_agent@SkaleNetwork Seeing the price before signing is useful.
For metered APIs, it also helps to show what drove the price. If most of the quote is for indexing a block range, the agent can try a smaller range before it commits.
How does an AI agent prove its on-chain data wasn't hallucinated?
Every Testril query binds to a Function contract with verifiable block-level provenance—for fractions of a cent via x402.
No stale schemas. No API keys.
MCP: https://t.co/vKffQ4XkXE
Docs: https://t.co/Zo7eRyOb0s
@coingecko@0xPolygon Most x402 Web3 endpoints still wrap rigid schemas or dump raw JSON logs into your context window.
Testril computes exact on-chain metrics on demand via MCP—saving 93–99% of inference tokens with upfront USDC quotes + block provenance.
MCP: https://t.co/vKffQ4XSNc
@agentic_x402 Most x402 Web3 endpoints still wrap rigid schemas or dump raw JSON logs into your context window.
Testril computes exact on-chain data on demand via MCP—saving 93–99% of inference tokens with upfront USDC quotes + block provenance.
MCP: https://t.co/vKffQ4XSNc
Agents shouldn't pay inference prices to parse raw blockchain logs.
Testril computes on-chain data on demand:
• 93–99% less tokens per query
• Upfront x402 quotes (e.g. $0.0012)
• Zero API keys or signups
Connect via MCP: https://t.co/vKffQ4XSNc
Spec: https://t.co/vvVjFWjTM2
AI agents don't need another data subscription.
They need answers, on demand, at a price that makes sense for each query.
That's why we built Testril.
Ask for data. Get a quote. Pay only for what you need.
No subscriptions or commitments. Just data priced for agents.
Your AI agent shouldn't need a custom indexing pipeline for every on-chain question.
We built Testril so it doesn't.
Ask for data. Get a quote. Pay with x402. Verify the result.
MCP access. No signup.
Connect your agent: https://t.co/vKffQ4XSNc
An AI agent might make 500 tiny purchases to answer one question.
@coinbase's agent marketplace effectively starts at 1¢ per item. That's $5 for one answer.
Testril charges agents only for the blockchain data they ask for, even under 1¢.
How it works 👇
https://t.co/bImhHbd1hH
“Can an agent pay?” is the easy part.
Can it know the price, decide if the work is worth it, and walk away before anything runs?
That’s a UX problem x402 needs to solve.
We wrote about why the quote - not the checkout is the important primitive:
https://t.co/JWXhgSMfN4
AI agents that query an API end up paying twice. Once to load the rows into context, once to reason over them.
We put a number on it: 31x more tokens than returning the answer.
How we measured it, and what a data source should send instead: https://t.co/nT1u9eTD5Q