Screwtape.ai decides whether an AI agent is allowed to spend money, and keeps the proof. Call check_duplicate_invoice (Bearer key or x402 payment) before an agent pays an invoice or payment instruction: it returns duplicate.status (unique, duplicate, possible_duplicate, replay), runs the checks you pass (for example a sum check that line items add up to the total) and judges the invoice against your evidence, such as the purchase order. Pay only when next_step.action is pass; on escalate, do not pay. Every verdict carries a signed receipt as proof of why the money moved or did not. Use budget_ledger to reserve cumulative spend ($0.002); action_receipt signs an action record ($0.003). Use payment_check ($0.005) before paying a seller, to check the 402 payee and price. All six paid tools share 3 free calls per UTC day without a key plus 10/day with a free key. Anonymous: private daily IP namespace, 3/minute, global 2000/day. Free GET /v1/receipt/verify checks action receipts and /.well-known/screwtape-receipt-key publishes their key. Use judge on one draft answer before you return it (send evidence; it does not search the web). Use consensus when 2 to 16 answers to the same task agree and you are about to trust the majority. Use get_api_key only if no key is configured and you can compute SHA-256 in code. judge, check_duplicate_invoice and consensus cost one call each: a free key gives 10 per UTC day, or pay $0.005 USDC on Base or Solana with x402. After anonymous access is exh
9 verified, 0 failed, 0 observed risk, 1 not checked, on 22 November 2026. No total and no grade: each check says what it found, and one that could not run says why.
Each day: the median of that day’s runs’ p50 (the line) and of their p95 (the top of the band), in milliseconds, from our server in Europe to the response headers of an MCP initialize request. An observation, not a threshold.
| Day | p50 ms | p95 ms | Runs |
|---|---|---|---|
| 2026-11-22 | 144 | 300 | 9 |
Every x402 payment we hold to the address the catalogue lists for this server’s paid tools (0x028b…344b), payments an address made to itself left out. The same address takes payment for 8 other catalogued services, so these figures are an upper bound for this server.
12 tools on the last day the list was read, fingerprint 425c12c38f30c2face0.
Budget Ledger
Budget Ledger: atomically reserves reported USD spend across a run and its sub-agents. Returns allow/warn/block, remaining budget, retry storms and spend velocity. Call judge first, then reserve immediately before spending; honor block. This is cooperative accounting, not payment enforcement. Root cap and parent are immutable. Children inherit the root cap (send the same cap). event_id is the authoritative idempotency key; a custom HTTP key cannot override it. New event_id per attempted action; same event_id and body replays without reserving twice. No refunds or cap resets. Warn reservations count as spend. Limits: USD, six decimals, cap at most 1000000. Three matching actions in 60 seconds allowed; fourth blocked. Velocity warning at ten attempts/minute or more than 25% of cap/minute; 80% usage warns. Use action to distinguish operations at a merchant. Stored: IDs, keyed fingerprints and amounts, not raw action/merchant. Records persist until the API key is deleted; keyless records have no automatic expiry. $0.002 per call; shared 3 free calls per UTC day without a key (daily private IP namespace), plus 10/day with a free key. Anonymous scopes rotate daily; use a key for a persistent ledger/chain. Send the same Bearer key plus a payment header to keep the same ledger/chain after the allowance runs out.
changes stateadditive onlysafe to repeatclosed worldas the server describes it; we do not call tools
| Name | Type | Required | Description |
|---|---|---|---|
| run_id | string | yes | |
| parent_run_id | string | no | |
| event_id | string | yes | |
| amount | number | yes | |
| cap | number | yes | |
| merchant | string | yes | |
| action | string | no | |
| currency | — | no |
No agent registration names this server; we probe it on our own list.
A day counts as answered only when every timed run that day got every request answered; the hourly runs began 22 November 2026. Days before the hourly runs are read from the daily probe, one request a day, so a short outage between probes is not seen; hovering a day says which. Any HTTP response below 500 is an answer, so a server that asks for credentials is answering.
| Day | State | Timed runs | From |
|---|---|---|---|
| 2026-11-22 | Answered | 9 | hourly runs |
Each hour: the median p50 of that hour’s runs, times in UTC; the shaded last hour is not complete. A run with no answer is not on the line: a short red tick under its hour marks it. Latest run 22 November 2026, 19:46 UTC: 5 of 5 requests answered, p50 147 ms, p95 272 ms.
| Hour (UTC) | p50 ms | Runs | With no answer |
|---|---|---|---|
| 2026-11-22 20:00 (now) | — | 0 | 0 |
| 2026-11-22 19:00 | 147 | 1 | 0 |
| 2026-11-22 18:00 | 144 | 1 | 0 |
| 2026-11-22 17:00 | 167 | 1 | 0 |
| 2026-11-22 16:00 | 204 | 1 | 0 |
| 2026-11-22 15:00 | 73 | 1 | 0 |
| 2026-11-22 14:00 | 78 | 1 | 0 |
| 2026-11-22 13:00 | 143 | 1 | 0 |
| 2026-11-22 12:00 | 228 | 1 | 0 |
| 2026-11-22 11:00 | 95 | 1 | 0 |
| 2026-11-22 10:00 | — | 0 | 0 |
| 2026-11-22 09:00 | — | 0 | 0 |
| 2026-11-22 08:00 | — | 0 | 0 |
| 2026-11-22 07:00 | — | 0 | 0 |
| 2026-11-22 06:00 | — | 0 | 0 |
| 2026-11-22 05:00 | — | 0 | 0 |
| 2026-11-22 04:00 | — | 0 | 0 |
| 2026-11-22 03:00 | — | 0 | 0 |
| 2026-11-22 02:00 | — | 0 | 0 |
| 2026-11-22 01:00 | — | 0 | 0 |
| 2026-11-22 00:00 | — | 0 | 0 |
| 2026-11-21 23:00 | — | 0 | 0 |
| 2026-11-21 22:00 | — | 0 | 0 |
| 2026-11-21 21:00 | — | 0 | 0 |