forge

Why we included this project

A local model gives you privacy and low cost, but the tool calling often breaks in ways that make it hard to trust. Forge wraps that loop and adds guardrails: it rescues malformed calls, nudges retries, and validates responses, and the project's own eval suite credits them with lifting a small 8B model from single digits to 84% success. The proxy is the quickest way to use it, since opencode, Continue, aider, and Claude Code can point at forge and keep working as if nothing changed. There's also a WorkflowRunner for structured agent loops with optional required steps and terminal tools, and a middleware mode for loops you already run. It works with Ollama, llama.cpp, Llamafile, vLLM, and Anthropic backends, which makes it a practical middle ground for teams that want self-hosting without hand-rolling the parsing.