Forge is an open-source Python framework with 2,200+ stars on GitHub that makes self-hosted LLM tool-calling reliable, adding guardrails, retries and workflow constraints around models you run yourself. Note: this Forge is a distinct open-source project by developer antoinezambelli — it is not the other product named Forge already listed on Tooloso.
Small local models are notoriously sloppy tool callers: malformed arguments, skipped steps, calls in the wrong order. Forge, published on PyPI as forge-guardrails and updated very recently in 2026, sits inside a single agentic loop and fixes that layer. You hand it a set of tools and let the model call them freely, while forge’s guardrails — rescue parsing of malformed output, retry nudges and response validation — apply automatically. When you need structure, it is opt-in: required_steps, prerequisites and a terminal_tool constrain the loop without turning forge into an orchestrator. The effect is measurable: the project reports an 8B local model jumping from single digits to 84% on its 26-scenario evaluation suite, and even lifting a frontier model from 85% to 98% on the same workload. It works with llama.cpp, llamafile and Ollama setups.
Forge is completely free and open source under the MIT license. It runs against models you host yourself, so there is no subscription, hosted tier or per-call fee for the framework.
Developers building agents on local or self-hosted LLMs who are tired of flaky tool calls and want reliability engineering around the model instead of a bigger model bill.