A slow MCP server does not error — it silently vanishes from tool discovery when its cold start exceeds the client's handshake budget (~5s add / ~10s connect / ~25–30s tools/list), while its /healthz still answers 200. First-Call Witness runs the real MCP initialize + tools/list handshake from outside, grades each phase against a versioned per-client budget table (budgets-v1-2026-10-20), and issues a signed, expiring, budget-table-versioned time-to-first-tool receipt a registry or buyer can verify. Optional handshake keep-warm completes the full handshake — never a ping — to hold a scale-to-zero server inside budget.
One-off witness check $0.25/check · Scheduled checks metered at $0.25/probe · Handshake keep-warm $9.99/mo per server (x402; first check free, one time per operator).
Service: https://textforagents.com/products/first-call-latency-witness/ · Docs, budgets, terms and privacy are served from that base without auth.
Limitations, stated plainly: localhost measurement omits TLS, remote RTT and egress — real remote probe cost is measured against real external MCP targets; live client budget re-measurement is out of scope — the versioned budget table and 24h receipt expiry are the drift mechanism; results rest on a simulated scale-to-zero server speaking the real protocol, and real public MCP targets are additionally witnessed end to end.
Support: support@textforagents.com — verdict/receipt questions, budget-table questions, and keep-warm issues.