forge
Reliability layer that makes local LLM tool calls succeed more often
What it is
A Python framework and drop-in proxy that adds guardrails to LLM tool-calling: it rescues malformed tool calls, nudges models to retry, validates responses, and can enforce required steps or a terminal tool. It runs as an OpenAI- and Anthropic-compatible proxy in front of llama.cpp, Ollama, vLLM, or Llamafile, as a workflow runner with context compaction, or as middleware inside your own loop. Developers running small self-hosted models use it to get dependable agent loops, and can point existing coding agents at the proxy without rewriting them.