ypollak2/llm-router

v7.1.0 — Quota-Balanced Routing & Cross-Subscription Load Balancing (2026-04-21)

New feature: Automatically balance usage across Claude, Gemini CLI, and Codex subscriptions.

Added

• QUOTA_BALANCED Routing Profile — Dynamically reorder chains to balance quota consumption across three subscription providers

  • Monitors real-time pressure: Claude (session/weekly limits), Gemini CLI (daily), Codex (daily)
  • Within ±10% band → use free-first tiebreak order (codex → gemini_cli → claude)
  • Imbalance > ±10% → route to least-used provider first
  • Prevents one subscription from being exhausted while others remain underutilized

• llm_quota_status MCP Tool — Real-time visibility into subscription quota balance

  • Shows usage % for each provider
  • Route priority recommendations
  • Time to next reset (UTC midnight for Gemini CLI/Codex, custom for Claude)
  • Balance metrics and reordering decisions

• Codex Daily Quota Tracking — Local counter for OpenAI free tier (1000 req/day)

  • Persisted in ~/.llm-router/codex_quota.json
  • Auto-resets at UTC midnight
  • Integrated with quota-balance calculations

• Gemini CLI Quota Recording — Increment counter on successful requests

  • Alias record_gemini_request() for router integration
  • Complements existing get_gemini_pressure() monitoring

Configuration

# Use QUOTA_BALANCED to automatically balance subscriptions
llm_router_profile = "quota_balanced"

# Or configure via env:
export LLM_ROUTER_PROFILE=quota_balanced

# Codex daily limit (default 1000 for free tier):
export CODEX_DAILY_LIMIT=1000

Technical

• New module: src/llm_router/quota_balance.py (quota tracking + chain reordering)
• Router integration: _build_and_filter_chain() applies quota-aware reordering when profile == QUOTA_BALANCED
• Request recording: Added after successful Codex/Gemini CLI calls in _dispatch_model_loop()
• Type definition: Added QUOTA_BALANCED = "quota_balanced" to RoutingProfile enum

Performance

• Quota checks are async and cached per request