A universal, zero-dependency model-routing shunt. Delegates massive codebase reading and boilerplate generation to fast, economical worker models (Gemini Flash, Groq, Ollama) — cutting token burn by up to 90% while preserving context.
A token bill audit: what your agent was paying, and what it pays now.
Inspired by Spotify Engineering's internal routing pattern, redesigned as a universal open-source tool.
The primary frontier agent reads raw files directly into its expensive context window.
Heavy files stream into a fast worker model. Only concise bullets with exact line numbers return.
Estimate how much money and context tokens Model-Shunt saves for your daily coding workflow.
These runs called the worker directly on real files. Factual claims (npm d3, Kubernetes OpenAPI) were checked against the source. On mypy, 6 of 13 citations name the symbol on that line.
Binary files (PDF, images, wheels) are rejected before any API call · Full methodology: docs/BENCHMARK.md
Copy and paste the configuration into your preferred coding agent.
Every detail designed to remove friction, prevent bugs, and safeguard your context window.
Written in pure Python 3 standard library. No pip install, no virtualenv, and no npm. Works instantly anywhere Python 3.9+ is installed.
Bypasses the 128 KB CLI buffer limit in Linux by streaming file contents over stdin. Process hundreds of files in one pass without shell crashes.
Exposes get_available_models to query your provider's active models in real time, auto-selecting the best model for reading vs writing.
Automatically scans initial byte blocks for null bytes (\x00) to reject images, PDFs, or compiled objects before sending them to the LLM.
Every file block has each line prefixed with its 1-based number. Worker models cite exact line ranges with zero guessing or hallucination.
Automatic retry loops for HTTP 429 rate limits and 503 gateway timeouts, ensuring long-running agent workflows don't fail unexpectedly.