Introduction
Harnez is a minimal harness for coding agents. It ships a server process that runs the agent loop, and a terminal UI to drive it. That's the whole surface area.
It is model-agnostic. Plug in any provider through the underlying agent core. Otherwise, it stays out of your way: no dashboard, no plugin marketplace, and no config you didn't ask for.
bun packages/server/src/index.tsWhy a harness, not a framework
Harnez is opinionated about one workflow: terminal-first coding agents. It doesn't try to be a general extension platform. There's no plugin marketplace and no public SDK. Good internal modularity is a goal; public extensibility isn't.
Where to go next
Start with Installation to get the binary running, then read Architecture to see how the pieces fit together.
Design principles
- Coding-first, not coding-bound. Software development gets the strongest UX and tooling, but the runtime doesn't assume it. Research, document analysis, and other agentic workflows work too.
- Model-agnostic. Models are runtime dependencies, not architectural foundations. Different agents in the same session can use different models.
- Trusted execution. Harnez doesn't gate ordinary reads, writes, edits, or shell commands behind approval prompts. Control comes from visibility, steering, and interruption instead.
- Context is scarce; execution state is not. Large tool output is externalized rather than stuffed into the model's context window.