TokenMiser v0.5.0
First feature release: TokenMiser is now a full Rust LLM cost-routing proxy — a 9-crate workspace built on Pingora.
Highlights
- Drop-in OpenAI-compatible proxy —
/v1/chat/completions, streaming (SSE) and non-streaming, with per-response audit headers (x-tokenmiser-cache,x-tokenmiser-tier,x-tokenmiser-difficulty,x-tokenmiser-routed-to) - Tiered router — Tier 0 heuristics → Tier 1 semantic classifier → Tier 2 speculative cascade, plus a Rhai policy DSL with
tokenmiser policy testreplay - Dual-layer semantic cache — L1 exact-match LRU + L2 semantic (bge-small embeddings, precision-tuned 0.87 cosine threshold)
- Provider adapters — OpenAI, Anthropic, Ollama behind one trait; local-first: auto-detects a running Ollama and routes easy traffic there with zero API keys and zero config
- Cost ledger + live dashboard — real-time USD spent/saved from a canonical
pricing.json; htmx dashboard at/and JSON/stats - Quality judge — shadow-mode A/B with LLM-as-judge win-rate auto-gate
- MCP budget gateway — per-agent/per-tool spend caps via
/v1/mcp/*
No telemetry, ever. Full details in CHANGELOG.md.
Install (macOS arm64)
curl -sL https://github.com/openintelligence-labs/tokenmiser/releases/download/v0.5.0/tokenmiser-v0.5.0-aarch64-apple-darwin.tar.gz | tar xz
./tokenmiser-v0.5.0-aarch64-apple-darwin/tokenmiserOther platforms: cargo build --release from source.
SHA-256 (tokenmiser-v0.5.0-aarch64-apple-darwin.tar.gz): 888c0e8e861ea7f142d0f72a25a0b73147ccc11a04c782388871183792a21a3e