The counterparty-risk layer for agent payments.

npm i @forum-labs/payfetch
claude mcp add payfetch \
  --env PAYFETCH_PRIVATE_KEY=0x... \
  --env PAYFETCH_TEST_MODE=1 \
  -- npx -y -p @forum-labs/payfetch payfetch-mcp

Claude Code agents are already meeting HTTP 402 paywalls today, and the MCP payments extension is moving toward a release candidate. The safe time to give an agent a budget is before it needs one. Payment-path package: read what payfetch does not protect against.

As of July 6, 2026, of the 24,036 endpoints we have probed across the public x402 directories, 77% did not answer our probes with a valid payment challenge — an upper bound on the dead shelf, not a body count: some of those endpoints are genuinely dead, others answer a different HTTP method or x402 dialect than our probe spoke. We have upgraded the probe to try both (see the trust methodology) and are re-measuring; the corrected breakdown will be published here with raw counts. Pay-to addresses changed hands 872 times, including a handful of endpoints that rotate their payee on nearly every probe. None of this is visible to an agent at the moment it pays. That is why the record exists, and why it is append-only.

Agent commerce is commerce without recourse. The x402 protocol turns any HTTP 402 into a machine-payable endpoint. An autonomous buyer that pays the wrong one gets no chargeback, no contract, and no brand to remember. Its only protection is knowing the counterparty before it pays. Forum Labs is one network with three surfaces. A paying-fetch client bounds what an agent can spend. Two independent checks, endpoint reliability history and token-deployer history, let it vet the counterparty before the money moves.

One network, three surfaces

The client holds the seat. Payfetch is where the payment actually happens: hard caps, a dedicated small-balance wallet, an approval threshold you set, and a local receipt for every payment. Free and non-custodial, it stands on its own with no trust data at all.

The record is append-only. Our prober continuously watches the public x402 universe and writes what it sees into a timestamped record that is never silently edited: outages, payee and custody changes, price drift, deaths and resurrections. Token-deployer launch histories are archived the same way. History like this cannot be reconstructed later; it only exists because something was watching when it happened.

The verdicts are the surface. The Trust Score and the Safety Screen turn that record into per-call answers an agent can consult before it pays, and every score ships with the raw counts behind it, so you can re-judge our verdicts yourself.

Each piece is useful alone. Together they are one loop: the client can check the record before it pays, and the record deepens as the network watches.

Independent, non-custodial, and on the buyer's side. We don't run a payment rail, don't sell directory placement, don't take custody of funds, and don't accept payment from any endpoint or token we measure. Every score ships with the raw counts behind it. Metrics are mechanical and informational, not financial advice.

For endpoint operators

Forum Labs · Payfetch · Trust Score · Token Safety · Methodology · GitHub · ops@forum-labs.com · @shopforumlabs
© 2026 Forum Labs. Metrics and tools are informational and mechanical. Not financial advice.