Skip to main content
A claim about a skill has a whole life: it is opened with money behind it, an agent buys and tests the skill, and if the skill misbehaves the money moves. Here is that loop, start to finish.
Before any of it, a skill is listed. Open Register a skill in the app, pick one of the bundled examples or paste your own small Node program, and press Register. The gateway hashes the source into the skill’s id, pins the bundle on Hedera and announces it. Listing costs nothing; money enters only with a claim.
The five-step claim lifecycle

How a claim lives and dies: staked on Arc, bought over x402, probed in the sandbox, disputed with a bond, settled by a re-run.

1

Claim — lock the money on Arc

A participant picks a skill and one of six testable rules (called predicates):
  • NO_ENV_READ_OUTSIDE — reads no environment variable outside an allowlist.
  • NO_NET_EGRESS_OUTSIDE — opens no network connection outside an allowlist.
  • NO_FS_READ_OUTSIDE — reads no file outside the skill’s own folder.
  • NO_FS_WRITE_OUTSIDE — writes no file outside the skill’s own folder.
  • NO_CHILD_PROCESS — never starts a subprocess or worker.
  • NO_DYNAMIC_CODE — never runs eval, new Function or vm code.
The last two have no allowlist on purpose: the sandbox refuses every subprocess and every eval, because an allowed one would run outside the sandbox and take the proof with it. Those claims say simply that the skill never tries.They prove they are a real, distinct human with World Selfie Check, then lock USDC in the SigilStake contract on Arc. The claim is announced on a Hedera registry topic and gets its own audit topic.
2

Consume — an agent pays to use the skill

An agent reads the public list of skills (GET /skills), applies a simple, printed policy (only skills with real money staked on Arc are worth trusting), and asks the gateway for the source.Without a licence, the gateway answers 402 Payment Required. The agent pays a tiny per-kilobyte fee over x402, the Blocky402 facilitator settles it on Hedera, and the gateway hands over the source and mints a licence NFT to the payer. A person can do the same with one command: pnpm buy <skill>.
3

Probe — run the skill in a sealed sandbox

The agent runs the skill in a deterministic sandbox, with a normal input and a malformed input, because error paths are where secrets leak. The sandbox denies everything by default: fake “canary” secrets are never readable, files outside the skill’s folder fail, and every network call is refused unless the claim’s allowlist permits it. Every blocked attempt is logged, and the whole run is hashed into a trace.
4

Dispute — put a bond behind the evidence

If the skill broke the rule, the agent posts a counter-bond (at least 25% of the stake) and the trace hash to SigilStake.dispute() from its own Circle Agent Stack wallet, and uploads the full evidence bundle to Hedera. No human is asked to judge.
5

Settle — the money moves

An independent verifier re-runs the exact same bundle. If it reproduces the violation, the disputer wins; if it does not, the original staker keeps the pot. SigilStake.resolve() pays out on Arc, and the outcome is written to Hedera. A claim that nobody breaks within the cooldown can be withdrawn with the stake intact.

​
The rule that makes it real

The verifier’s decision is a single line, and it trusts no opinion:
Anyone can download the evidence bundle from Hedera and run it themselves to check the outcome. That is the whole point: the truth is reproducible, not asserted.

​
Try the buyer’s side in 30 seconds

Every skill page has a “Use this skill” section. Press Request the source and you will see the exact 402 paywall an agent gets. A browser cannot pay it (paying needs a signed Hedera transaction), so to actually fetch a skill you run one command:
That pays the fee once, saves the skill’s files locally, and gives your account a licence so every later download is free.