antoinezambelli/forge at 1.4K Stars With +449 Today: Reliability Layer Makes Self-Hosted LLMs Hit 86.5% on Tool-Calling Benchmarks

GitHub·medium signal

Forge adds guardrails — parsing rescue, retry nudges, and step enforcement — to self-hosted LLM tool-calling workflows, achieving 86.5% accuracy on a 26-scenario evaluation suite using Ministral-3 8B. Supports Ollama, llama-server, Llamafile, and Anthropic backends. Three usage modes: WorkflowRunner for structured agent loops, guardrails middleware, and a proxy server intercepting requests. Accompanied by an academic paper.