AgentTether

AI agents buy blockchain event watches with x402: the agent signs one Permit2 voucher (worst case = the quoted ceiling), a Substreams stream meters the blocks actually processed, and settlement is pro-rata — exactly the blocks billed, delivered to the agent's webhook only after the settlement receipt confirms.

Recent requests

The most recent watches (watch parameters only — wallets are never shown). Click any to see its flow: every step with ticks and timestamps.

Ask your agent

Give your coding agent (opencode, Claude Code, …) a prompt like:

"Using the AgentTether API at http://localhost:8080, watch Ethereum mainnet USDC
for any transfer ≥ 1,000 USDC for the next 60 seconds. Pay for it yourself with x402
(the first run does a one-time Permit2 approval — handled automatically by the agent),
then report what arrived."

"Using the AgentTether API at http://localhost:8080, watch our project's test wallet
0xd8dA6BF26964aF9D7eEd9e03E53415D37aA96045 for an incoming MATIC transfer of at
least 10 MATIC for the next hour. Pay for it yourself with x402. Once the account
has enough MATIC, run our project's verification scripts and if those pass, push
our code."

One constraint for now: watches observe ERC-20 Transfer events only, so a watch — asset-based or wallet-based — has to be about transactions involving an ERC-20 token.

The 202 response carries a report_url back to your agent — that's a link into this site's flow-chart rendering of the whole process.

The agents driving this are plain x402 clients — the API is agent-agnostic by design. Our own test agents use LangGraph for one reason: durable interrupts. The graph parks at wait_for_webhook with zero compute while the watch runs, and the backend's webhook resumes it — the pause is the point.