Skill119 repo starsupdated today

create-harness

The create-harness skill scaffolds a branded AI agent harness with its own npx CLI, MCP server registration, memory backend, and npm-publishable package structure. Use this when users request custom agent harnesses for vertical domains like legal, trading, or support, or when they want to build Claude Code plugins, Codex skills, pi.dev extensions, or Hermes runtime configurations with their own agents and tools bundled together.

View source Repository: agent-harness-generator

Install in Claude Code

Copy

git clone --depth 1 https://github.com/ruvnet/agent-harness-generator /tmp/create-harness && cp -r /tmp/create-harness/.claude-plugin/skills/create-harness ~/.claude/skills/create-harness

Then start a new Claude Code session; the skill loads automatically.

Definition

SKILL.md

# create-harness

This skill scaffolds an AI agent harness — your own focused, branded harness with its own `npx <name>` CLI, MCP server registration, memory namespace, learning loop, and marketplace identity.

## When to use this skill

Use this skill when the user wants any of:

- A custom MCP-server-backed AI assistant (legal, trading, support, ops, …)
- A Claude Code plugin that bundles their own agents/skills/prompts
- A Codex skill that wraps a kernel + tools
- A pi.dev extension or Hermes agent runtime config
- A standalone npm package they can `npm publish` under their own scope

## How to invoke

```
/create-harness
```

The skill asks for:

1. **Harness name** (kebab-case, e.g. `legal-redline`)
2. **Description** (one line)
3. **Host(s)** — Claude Code, Codex, pi.dev, Hermes (multi-select)
4. **Template** — `minimal` (default), `vertical:trading`, `vertical:support`, `vertical:devops`, `vertical:legal`, `vertical:research`, or `eject-from-ruflo`
5. **Memory backend** — AgentDB (default), SQLite-only, in-memory
6. **Routing strategy** — 3-tier (default) or single-tier
7. **Marketplace** — independent (no ruflo branding) or powered-by

## Under the hood

Calls `npx create-agent-harness <name>` with the user's choices, copies the chosen template into the current directory, runs the post-generation smoke test, and emits a `.harness/manifest.json` so future updates can `harness upgrade` cleanly (copier-style regenerate-diff-merge).

## Outputs

| File | Purpose |
|---|---|
| `package.json` | The harness's npm package, ready to `npm publish` |
| `bin/<name>.js` | The CLI binary your users invoke |
| `.claude/settings.json` (if Claude Code host) | MCP + hooks wiring |
| `.codex/config.toml` (if Codex host) | MCP table in TOML |
| `AGENTS.md` + `SYSTEM.md` (if pi.dev host) | Pi extension instructions |
| `cli-config.yaml` (if Hermes host) | Hermes runtime config |
| `.harness/manifest.json` | Generator state — drives drift detection (ADR-008) |
| `witness.json` | Ed25519-signed provenance manifest (ADR-011) |

## After scaffolding

The user can:

```bash
cd <name>
npm install
npm test
npm publish              # ship their harness to npm
# their users: npx <name> init
```

## Notes for the model

- This is a deterministic operation — defer to `create-agent-harness` for the actual file creation, do NOT generate harness file contents yourself
- The witness manifest is required — do not skip it
- If the user asks for "ruflo for X" or "my own ruflo", recommend `vertical:<X>` templates first and offer `minimal` as the fallback
- Multi-host is supported but increases the harness's surface area — recommend single-host for first-time users unless they explicitly need multi-host