E

EvalCore

Records real AI application behavior and replays it offline in CI, helping you find regressions caused by changes before users do.

free
Last Updated 9/11/2026

About EvalCore

EvalCore is an Apache-2.0 open-source AI application evaluation tool: after changing prompts, switching models, or upgrading dependencies, it records real AI behavior and compares it with a baseline to find regressions before users do. Any service that supports HTTP or shell can be an evaluation target, with no SDK and no dependency on a programming language. On the first run, it calls the real model and records requests and responses in a local SQLite cassette; thereafter, each PR replays it offline in CI, without an API key, so identical inputs produce identical conclusions, and with --baseline main it fails only on regressions. The tool is a single dependency-free binary with no telemetry, registration, or server required; the official page lists CI use as $0.

Best for

AI application developers
LLM application engineering teams
Engineers responsible for CI/CD quality gates

Best use cases

Check whether AI application behavior has regressed after changing a prompt or model

Set up an offline regression gate for PRs so CI fails only when degradation appears

Build evaluation datasets for services accessed over HTTP or shell

Track token usage and cost for each evaluation run

Features

Behavior recording

On the first run, it calls the real model, records every request and response in a local SQLite cassette, and summarizes token usage and cost.

Offline replay

In CI, it replays the cassette without network access or an API key; identical inputs always produce identical conclusions.

Baseline gate

Compare against the baseline with --baseline main and fail only on regressions; CI uses the exit code as a gate.

Broad compatibility

Supports OpenAI, vLLM, Ollama, any OpenAI-compatible gateway, your own REST API, shell commands, and OTel / OpenInference tracing.

Pros

  • Single dependency-free binary usable with any language
  • Offline replay makes identical inputs produce identical conclusions, avoiding evaluation flakiness
  • Apache-2.0 open source; the official page lists CI use as $0
  • No telemetry, registration, or server required

FAQ

The official page labels it Apache-2.0 open-source software; CI use is $0, and it requires no registration, telemetry, or server.
On the first run, it calls the real model and records requests and responses in a local SQLite cassette; thereafter, each PR replays the cassette offline, without network access or an API key, and CI uses the exit code as a gate.

Tool Information

DeveloperEvalCore
Release Date9/11/2026
Pricingfree