harness-kit is a plugin collection for Claude Code. Each plugin packages a specific workflow — something like code review or codebase exploration — as a slash command you can install once and use in any project.
Your harness is the configuration that tells Claude how to work in a given session — files like CLAUDE.md and AGENT.md, plus any installed plugins. harness-kit helps you build and share that setup so it's not stuck on one machine.
The term is gaining traction — LangChain published "The Anatomy of an Agent Harness" and ships deepagents, which they explicitly call a harness.
You can. The issue is that good prompts tend to disappear — scattered across projects, left behind when you move to a new machine. Plugins give them a stable home with a version number.
A skill is a SKILL.md file — the workflow Claude reads when you type a slash command. A plugin is the package that wraps it: a directory with the skill, optional scripts, and a version. You install the plugin; Claude reads the skill.
See Plugins vs. Skills for the full breakdown.
A prompt tells Claude what to do. A SKILL.md specifies how: step ordering, input parsing, output format, error handling. That structure is what makes the same command produce consistent output every time.
The install system targets Claude Code's plugin marketplace. But SKILL.md files are plain markdown — any tool that reads prompt files can use them. VS Code Copilot reads CLAUDE.md natively via chat.useClaudeMdFile. See the Cross-Harness setup guide for Cursor, Windsurf, and others.
Plugins are plain markdown files. There are no binaries, no background processes, no network calls on install. Some plugins include shell scripts (like research's index rebuild) that only run when you invoke the skill. You can read every file before installing.
No. harness-kit is free and open source (Apache 2.0). All you need is Claude Code.
Run /harness-export to write a harness.yaml with your installed plugins. Commit it to your dotfiles or send it to a teammate. They run /harness-import harness.yaml and pick what they want from an interactive list — your setup is a starting point, not something they have to replicate exactly.
Yes. Plugins install to your harness globally, not per-project. research and explain work in any codebase without extra configuration. Project-specific stuff goes in CLAUDE.md.
Yes. A minimal plugin is about 10 lines of boilerplate plus a SKILL.md. See the Creating Plugins guide. You can also start with a bare skill file in ~/.claude/skills/ and promote it to a plugin when you're ready to share it.
The installed files are plain text — edit them directly. For something more lasting, fork the repo and point your marketplace at your fork. If you improve something, a PR is welcome.
MCP servers give your AI new tools — database access, web search, external APIs. harness-kit is about what to do with those tools: structured workflows with defined steps and outputs. They work at different levels and don't overlap.
A list of your installed plugins with their sources and versions. Export it with /harness-export, commit it, and restore it anywhere with /harness-import or harness-restore.sh. See harness.yaml.example for the format.
Not necessarily. Project-specific workflows belong in CLAUDE.md. If you find yourself copying the same prompt into every new project, that's probably a sign it would work better as a plugin.
They operate at different layers and don't overlap.
| Layer | What it solves | Example |
|---|---|---|
| Configuration | How is this agent set up? | harness-kit |
| Tool communication | How does the agent call tools? | MCP |
| Runtime communication | How do agents talk to each other? | A2A |
| Developer SDK | How do I build an agent? | Claude Agent SDK |
MCP gives your agent tools. A2A lets agents delegate tasks to each other at runtime. The Claude Agent SDK is for building agent-powered applications. harness-kit describes the complete environment, plugins, MCP servers, instructions, permissions, so you can reproduce it on another machine or share it with a teammate. They compose well: a harness.yaml can declare MCP servers, the agent runs on the Claude SDK, and could speak A2A to other agents.
See Where harness-kit fits for the full breakdown.
They work at different layers and complement each other.
deepagents is an agent harness framework — you build agents with it. It's Python, built on LangGraph, and opinionated about the execution environment.
harness-kit is a harness configuration framework — you configure any harness with it. It works across Claude Code, Cursor, and Copilot without coupling you to a specific runtime or language. Think of it as dotfiles that work across tools.
LangChain demonstrated that harness changes alone moved a coding agent from Top 30 to Top 5 on Terminal Bench 2.0 — evidence that the configuration layer is a first-class performance variable. In principle, you could use a harness-kit profile to configure a deepagents-based harness. They don't compete — one builds agents, the other makes agent configuration portable.
The Harness Protocol is an open specification for portable AI coding harness configuration. It defines a vendor-neutral harness.yaml format validated by JSON Schema. harness-kit is the reference implementation, the same way Claude Desktop implements MCP. Any tool that correctly validates and applies harness.yaml per the spec is a conformant implementation; harness-kit is not required.
The spec is Apache 2.0. See Harness Protocol for how it relates to harness-kit.