antoinezambelli/forge
A reliability layer for self-hosted LLM tool-calling that improves model performance on agentic workflows with rescue parsing, retry nudges, and response validation.

Forge provides guardrails that make LLM tool-calling more reliable without requiring structured workflow definitions. It sits inside a single agentic loop and improves tool call accuracy through parsing rescue, retry nudges, and response validation. The framework includes a 26-scenario evaluation suite to measure performance, and can be deployed as a proxy server speaking OpenAI chat-completions and Anthropic Messages APIs to integrate with existing tools like Claude Code, opencode, and aider.