EvalCore

EvalCore is an open-source developer tool that lets AI engineering teams know when a change to a prompt, model, or dependency makes their LLM-powered app worse. Instead of forcing you into a proprietary SDK or test harness, it wraps around anything that speaks HTTP or shell. You describe an evaluation as a YAML file plus a JSONL dataset, point it at your target, then stack scorers on top to define what "good" means. The core innovation is the cassette: on the first live run, EvalCore calls the real model and records every request and response into a local SQLite cache keyed by a hash of the canonical request. That cassette can be committed to your repository. In CI, EvalCore replays the recording entirely offline with zero network calls, zero API keys, and zero cost, producing deterministic verdicts. The `--baseline main` flag runs a comparison against main and exits nonzero only on regressions, making it a natural pre-merge gate. The tool is designed to be lightweight and language-agnostic. It ships as a single dependency-free binary available via `cargo install` or prebuilt binaries for macOS, Linux, and CI runners. It works with OpenAI-compatible APIs, vLLM, Ollama, REST endpoints, shell commands, and OTel/OpenInference traces. With Apache-2.0 licensing, no server, no signup, and no telemetry, EvalCore positions itself as a simple, auditable way to ship AI changes confidently.