trystub
A drop-in budget gate for AI agents that spend real money: paid APIs, x402 micropayments, LLM tokens. One enforceable budget in front of the whole fleet, with an exact double-entry record of every agent's spend. Uses a reserve → pay → settle flow so irreversible payments fire exactly once, even under retries or crashes.
Problem
Autonomous agents spend real money on paid APIs, x402 micropayments, and LLM tokens, and nothing enforces a ceiling. A retry loop or a runaway sub-agent can burn a budget before any human notices, and payments are irreversible once fired.
Approach
- Put one enforceable budget gate in front of the whole fleet instead of per-agent limits that drift apart.
- Designed a reserve → pay → settle flow so irreversible payments fire exactly once, even under retries or crashes.
- Built meterLLM to price OpenAI- and Anthropic-shaped usage without shipping stale price tables.
- Added idempotency keys so retries are safe across the fleet, plus operator sweep and shard provisioning for sustained throughput.
Outcome
A drop-in npm package (zero dependencies) that gives operators a single, auditable budget boundary with an exact double-entry record of every agent's spend.
Architecture
const reservation = await stub.reserve({
agentId, amount, idempotencyKey,
});
if (!reservation.ok) throw new BudgetExceeded();
const receipt = await pay(reservation);
await stub.settle({
reservationId: reservation.id,
receipt, // safe to retry
});Highlights
Reserve → pay → settle exactly-once flow, so irreversible payments fire once even under retries or crashes.
meterLLM prices OpenAI- and Anthropic-shaped usage without shipping stale price tables.
Idempotency keys make retries safe across the whole agent fleet.
Operator sweep and shard provisioning for sustained throughput.
Stack
- TypeScript
- AWS
- Aurora DSQL
- x402
- Large Language Models