forge
A reliability layer that makes self-hosted LLM tool-calling more dependable, sitting inside an existing agent loop or coding harness as a proxy, a workflow runner, or middleware that validates responses and rescues malformed tool calls.
forge sits inside a single agentic loop and makes its tool calls reliable rather than orchestrating multiple agents or replacing a coding harness. Workflow structure such as required steps, prerequisites and a terminal tool is opt-in, but its guardrails, including rescue parsing of malformed tool calls, retry nudges and response validation, apply even with no required steps configured. In an internal evaluation suite, forge lifted an 8B local model's pass rate from single digits to 84%, and a stronger commercial model from 85% to 98%.
It can be used three ways: as a drop-in proxy server that speaks both the OpenAI chat-completions and Anthropic Messages APIs, sitting between an existing client such as Claude Code or another coding tool and a local model server; as a WorkflowRunner that manages a structured agent loop's full lifecycle, including a priority-queued SlotWorker for sharing one GPU inference slot across specialist workflows; or as composable guardrails middleware inside a user's own orchestration loop.
It supports OpenAI-compatible endpoints, Ollama, llama-server, Llamafile, vLLM and Anthropic as backends, and is released under the MIT license.