Forge, a Python framework from Texas Institutions' AI Director Antoine Zambelli, adds a reliability layer to self‑hosted large‑language‑model (LLM) tool‑calling. The system introduces guardrails—retry nudges, step enforcement, error recovery, and VRAM‑aware context budgets—that keep local 8B models running on consumer GPUs stable during multi‑step workflows.

On a single 8B model, Forge lifts accuracy from roughly 53 % to 99.3 % on multi‑step agentic tests without changing the weights. In benchmark runs, the same model outperforms the Claude Sonnet API baseline—dropping the performance gap between a $600 GPU setup and a frontier service to under one point and demonstrates efficiency and shows practical value.

Forge’s guardrail stack contains five toggleable layers; ablation tests show retry nudges alone can cause 24‑49 point drops when disabled, while error recovery accounts for a 10‑point fall across models. The framework also introduces a ToolResolutionError exception, forcing models to retry when a tool returns empty data instead of silently propagating garbage through subsequent steps.

Running Forge is straightforward: install via pip, choose a backend like llama‑server or Ollama, and launch the proxy to wrap any OpenAI‑compatible client. Adopters report that an 8B local model with Forge matches or surpasses cloud‑based APIs, while keeping inference costs below $20 per month on a single GPU. The project is open‑source on GitHub.