An AI layer on top of block explorers

Block explorers are very good at displaying data. They are not always good at explaining what happened.

A failed transaction can include hashes, logs, events, gas used, internal method calls and contract errors. For an experienced developer that is useful. For someone testing on a testnet for the first time, it is mostly noise.

The idea behind the AI Explorer on testnet.help is simple:

Ask a question in plain language. Get a useful answer backed by real on-chain data.

The problem with current explorers

When a transaction fails on Sepolia, Etherscan shows you this:

Status: Fail
Error: execution reverted

Sometimes there is a revert reason. Sometimes there is not. Either way, you are on your own to figure out what 0x08c379a0... means, whether the issue is an allowance, a reentrancy guard, a missing role, or a simple parameter error.

The cognitive work — connecting the raw data to an actionable explanation — falls entirely on the developer. That is fine if you have been doing this for years. It is a real barrier if you are learning.

What we built

The AI Explorer at /chat is a conversational interface that sits on top of public blockchain data. It supports six testnets today: Ethereum Sepolia, Base Sepolia, Polygon Amoy, BNB Testnet, Bitcoin Signet and Solana Devnet.

You can ask things like:

• “What is the balance of 0x1234… on Sepolia?”
• “Is this transaction confirmed on Bitcoin Signet?”
• “Which faucets work for Ethereum Sepolia right now?”
• “What testnets do you support?”

The assistant calls real APIs under the hood — public RPCs for EVM networks, Mempool.space for Bitcoin, the Solana RPC for Devnet — and returns a plain-language answer with explorer links when relevant.

Architecture: privacy first

The original design had the AI conversation running through our server. We scrapped that.

The current architecture keeps the API key entirely in the browser:

Browser ──(your API key)──▶ OpenAI / Anthropic
Browser ──(no key)────────▶ /api/blockchain/address/:network/:address
                             /api/blockchain/tx/:network/:hash
                             /api/blockchain/faucets/:network
                                    │
                                 server ──▶ Public RPCs / Mempool.space

The key never leaves your browser. Our server only handles the blockchain data lookups — stateless, keyless, no logs. You can use OpenAI or Anthropic depending on what you have.

This also means there are no rate limits on our end, no cost we absorb, and no account required. Bring your own key and the tool works.

Tool use under the hood

The assistant uses Claude or GPT-4o-mini with three tools defined:

• lookup_address — queries balance and activity for any address on any supported network
• lookup_transaction — fetches transaction status, block number, fees and explorer link
• list_faucets — returns faucet status from the live data in testnet.help

The model decides which tool to call based on your question, calls it, reads the result and formulates a response. You never see the JSON — just the answer.

This is the same pattern used in production AI agents. testnet.help is a concrete, minimal example of how tool use works in a real context.

What comes next

The current version answers questions. The next step is explaining things.

A transaction trace has a story: who called whom, which internal calls succeeded, where the revert happened and why. Right now cast run gives you that trace in the terminal. The next version of the AI Explorer should be able to read that trace and explain it in plain language — the same way the debug guide walks you through it manually, but automated.

Other directions worth exploring:

• Contract reader — paste an address, get a plain-language summary of what the contract does
• Revert decoder — paste raw revert data, get an explanation and a suggested fix
• Faucet status alerts — notify when a faucet changes status, via webhook or feed

The data infrastructure is already here. The faucet status, the network metadata, the RPC endpoints — all of it feeds into the same system. Adding a smarter reasoning layer on top is the natural next step.

If you want to try the current version, go to /chat. Bring a key from OpenAI or Anthropic, and start asking.