Rate Limiting Quotas and Budget Management

6.1 Rate Limiting

Configurable across every dimension the gateway sees: model, user, team, organization, API key, application, MCP server, and individual tool. Defaults: global (100/min per IP), per-user (20/min), agent proxy (60/min), MCP proxy (120/min), per-key (60/min), import (10/min). Per-tool and per-server quotas live on the MCP proxy (Section 14). Scalability: maxSockets=1000, maxConcurrent=500, queueSize=10000, queueTimeout=60s.

Fallback behaviour on limit breach: route to the next-priority provider (Section 8 health-priority fallback), return 429 with retry-after, or emit a webhook/email alert — selectable per scope.

6.2 Budget Management (4-Tier)

Organization ($10/day, $300/mo), Team ($50/day, $1500/mo), Tenant ($10/day, $300/mo), User API Key ($10/day, $100/mo). maxApiCallsPerMonth=0 is treated as unlimited (not zero-calls-allowed).

Hard enforcement, not just alerting: when a daily or monthly budget is exceeded, the gateway rejects the request with 402 Payment Required (configurable to 429) rather than letting the spend continue. This is the categorical difference from observation-only platforms that emit a Slack alert after the fact — AgentWatch can actually stop the spend.

Alert cadence: webhook + email alerts fire at 80% / 90% / 100% of each tier’s daily and monthly budget, with the recipient list configurable per tier.

Admin override with expiry: company admins can grant temporary budget overrides (extra headroom for a known launch event, a specific tenant doing migration, etc.) with an automatic expiry timestamp — no need to remember to revoke them.

Budget efficiency leaderboards: the analytics views rank teams/employees by tokens-per-dollar, by output quality vs. cost (when paired with downstream feedback), and by month-over-month spend trend — useful for FinOps reviews and for identifying which teams have adopted cost-aware patterns (semantic caching, prompt caching, cheaper models for non-critical workloads).

---