AuroraCore

OVERVIEW

AuroraCore is a coding agent written in Rust — a fork of OpenAI’s Codex that reimplements the agent harness with a different priority: getting strong performance out of cheap models rather than assuming the most expensive one is available. It runs against Kimi, DeepSeek, Claude, Qwen, and other providers.

THE PROBLEM

A coding agent’s results depend as much on its harness — how it frames the task, tools, and context — as on the model behind it. A harness tuned for one frontier model tends to waste tokens and lose accuracy when pointed at a cheaper one.

APPROACH

Harnesses are swappable at runtime through a `/harness` command, so the same agent can emulate the native, claude-code, kimi-code, or swe-agent shapes depending on what the model responds best to. The project deliberately builds on shared standards — AGENTS.md, `.agents/skills` directories, MCP, and the Agent Client Protocol — so that, as the README puts it, it "should fit into your existing agent setup instead of trapping it".

FEATURES

• Swap agent harnesses at runtime to match the provider in use
• Sandboxed command execution on macOS, Linux, and Windows
• Speaks the Agent Client Protocol and stays compatible with the Codex SDK
• Ships a QA skill for driving web and native applications

STACK

Rust / TypeScript / Bazel / MCP