Changelog
What's new and improved in RelayPlane.
v1.7.4March 2026
- •Per-agent cost tracking — agents identified by system prompt fingerprint; Agent Cost Breakdown table in dashboard
- •Request/response content logging — dashboard now shows system prompt preview, user message, and response preview per request
- •OAuth passthrough fix — correctly forwards user-agent and x-app headers for Claude Max subscription users (critical for OpenClaw integration)
- •Admin dashboard Phase 1 — new frontend for proxy management
- •Provider column in Model Breakdown — distinguishes anthropic vs openrouter for identically-named models
- •Agent column in Recent Runs table
- •Expandable rows that persist across the 5-second auto-refresh
- •Honest savings breakdown — routing savings (RelayPlane) vs cache savings (Anthropic) with tooltip explanation
- •Error detail capture — failed requests now show error message and HTTP status code
- •Wider dashboard layout (1600px) for dense data views
v1.7.0February 2026
- •Admin dashboard Phase 1 — new local dashboard at localhost:4100 with cost breakdown, savings analysis, provider health, and agent breakdown (1600px wide layout)
- •Budget enforcement — rolling daily/hourly/per-request spend limits backed by SQLite with <5ms hot-path checks; breach actions: block, warn, downgrade, or alert
- •Anomaly detection — real-time sliding window analysis catches velocity spikes, cost acceleration, repetition loops (possible agent hangs), and token explosions
- •Cost alerts & webhooks — threshold/anomaly/breach alerts with 5-min deduplication, webhook POST delivery, and SQLite history (up to 500 alerts)
- •Auto-downgrade — when budget hits configurable threshold (default 80%), proxy rewrites expensive models to cheaper alternatives; built-in mappings for Anthropic, OpenAI, and Google models; adds X-RelayPlane-Downgraded headers
- •Response cache — exact-match caching with SHA-256 keyed requests, in-memory LRU + gzipped disk persistence; aggressive mode available for broader matching
- •Rate limiting — per-workspace, per-model 1-minute sliding windows; Opus/o1 at 10 rpm, default at 60 rpm; returns 429 with Retry-After
- •Emergency kill endpoint — POST /admin/kill halts all request processing immediately
- •Startup config validation — proxy validates all config fields on boot before serving traffic
- •Clerk → Supabase auth migration in API layer
v1.6.0February 2026
- •Budget enforcement — per-session daily/hourly/per-request spend limits with block, warn, downgrade, or alert actions
- •Anomaly detection — real-time alerts on velocity spikes, cost acceleration, token explosions, and repetition loops
- •Cost alerts & webhooks — threshold alerts at configurable percentages with webhook delivery and alert history
- •Auto-downgrade — automatically routes to cheaper model when budget threshold is hit
- •Aggressive cache mode — maximize prompt cache hit rate with broader match and shorter TTL
- •Response cache — exact-match response caching with gzipped disk persistence to skip API calls entirely
- •Rate limiting + emergency kill endpoint
- •Startup config validation — proxy validates configuration on boot and reports errors clearly
v1.1.1February 2026
- •New default port (4100) to avoid dev tool conflicts
- •Added /health endpoint for proxy monitoring
- •Fixed quality routing mode (updated model IDs)
- •CLI now reads port from config file
- •Version command reports correct package version
v1.1.0January 2026
- •Relay Network for configurable model routing (auto, cost, quality, fast modes)
- •Cost calculator and live benchmarks pages
- •Supabase auth migration — faster, simpler sign-up flow
- •Interactive playground for testing routing strategies
- •Comprehensive provider support: Anthropic, OpenAI, Google, xAI
v1.0.0December 2025
- •Initial release of RelayPlane proxy
- •Drop-in replacement for OpenAI-compatible APIs
- •Automatic cost optimization with 40–60% savings
- •Free tier with 1,000 requests/day
- •Pro tier with unlimited requests and priority routing
For detailed release notes, see our GitHub repository.