Burnless: Reduce Multi-turn LLM API Costs by 90%
Test Burnless in your multi‑turn agent workflows to cut token costs.
Test Burnless in your multi‑turn agent workflows to cut token costs.
Summary
Burnless is an open protocol that turns quadratic multi‑turn LLM cost into linear by keeping only short capsules. It uses a shared prefix cache with Anthropic prompt caching (ttl 1h) and an 80‑character compressed capsule history to replace raw transcripts. Vendor‑agnostic orchestration lets you mix Claude, GPT, Gemini, Mistral, or local Llama via config.yaml tiers gold/silver/bronze. Benchmarks show a 10‑turn session with Claude 3 Opus costs $4.66 standalone, $0.65 with cache, and $0.45 with Burnless, a 90.3% reduction. The tool is installable via pip and supports any CLI already installed. Burnless is MIT licensed and aims to reduce token consumption for developers building agents.
Key changes
- Transforms multi‑turn agent cost from Θ(N²) to Θ(N) by keeping only short capsules.
- Implements shared prefix cache using Anthropic prompt caching with ttl 1h to avoid re‑billing large system prompts.
- Uses capsule history of ~80‑character compressed turns to replace raw transcripts.
- Supports vendor‑agnostic orchestration with tiers gold/silver/bronze defined in config.yaml.
- Benchmarks show 90.3% cheaper than Claude 3 Opus on a 10‑turn session (standalone $4.66 → Burnless $0.45).
- Allows mixing any CLI (Claude, GPT, Gemini, Mistral, local Llama) for brain and workers.
- Provides local Ollama execution for zero marginal cost on cheap stages.
- Installable via pip with `pip install burnless` and setup via `burnless setup`.