Agent Preflight
Decide before you spend a tool call.
Agent Preflight evaluates a proposed tool or API call before execution and returns an actionable decision so agents can avoid invalid, redundant, unsafe-to-retry, or unnecessarily expensive calls.
It validates structured search/browse tool intents against budget and history signals. It never calls the destination provider for you.
When to call it
- Before a metered search or browse tool call in a budgeted agent loop.
- When you have recent tool history and want duplicate / retry detection.
- When remaining budget may not cover the estimated external cost.
What it evaluates
- Tool provider/operation and request shape (supported:
search:web_search,browse:fetch). - Estimated external and model costs vs remaining budget.
- Optional history fingerprints for reuse / terminal retry caps.
- Optional capabilities such as reuse and scope reduction.
Decisions
allow— Proceed with the proposed tool/API call.block— Do not execute; the call looks invalid, unsupported, or proven wasteful.reuse— Skip a new call and reuse a recent successful result when available.defer— Do not execute now; usually budget-infeasible or not worth spending yet.advise— Optional narrowing advice; not a hard stop. Not billed as an actionable decision fee in economics.
Production endpoint and price
POST https://api.agent-preflight.com/v0.2/preflight/tool-call
$0.001 USDC per HTTP POST via x402 on Base mainnet (eip155:8453) when payment is configured (production is configured for paid collection). Unpaid requests receive 402 with a PAYMENT-REQUIRED header.
POST /check-request and MCP /mcp remain unpaid. A billable HTTP POST is separate from whether a decision is economically “actionable.”
Privacy
Analytics correlate runs using one-way hashes of run and step IDs. Request bodies, schemas, URLs, and raw run/step identifiers are not written to Analytics Engine points.