MetaHarness × Claude Code workspace + plugin

A one-command scaffold that drops a ready-to-edit Claude Code workspace into a new directory: CLAUDE.md, a .claude/ settings tree, MCP server stubs, and a minimal plugin layout you can load with claude -p --plugin-dir. This is the smallest useful starting point for building a Claude Code-hosted agent — not a multi-agent vertical, not a runtime, and not a chat UI. It's the workspace you wish claude init produced.

Use this when you want to ship a Claude Code agent that has opinions: custom slash commands, scoped tool permissions, a project-pinned MCP config, and a plugin folder you can iterate on without polluting your global ~/.claude config.

Quickstart

npx @metaharness/claude-code@latest my-bot
cd my-bot && npm install && npx harness doctor

harness doctor validates that Node is recent enough, that the .claude/ tree is well-formed, that the plugin manifest parses, and that any declared MCP servers resolve. From here you open the folder in your editor and start editing CLAUDE.md plus the files under .claude/.

What you get

• CLAUDE.md — project-level system prompt with behavioral rules, file-org conventions, and a starter agent routing table.
• .claude/settings.json — host-specific Claude Code config: permission allowlist, default model, hook stubs, and MCP server registrations.
• .claude/commands/ — example custom slash commands (/plan, /review) you can rename and extend.
• .claude/agents/ — one minimal subagent definition you can clone for coder/tester/reviewer roles.
• plugin.json + bin/ — a loadable Claude Code plugin manifest so claude -p --plugin-dir . picks it up.
• mcp.json — pinned MCP server config (stdio transport, no network defaults).
• package.json with a harness script bound to the metaharness CLI for doctor / validate / upgrade.

Advanced

# Sanity-check the scaffold
npx harness doctor
# → node: ok (v20.x)
# → .claude/settings.json: ok
# → plugin.json: ok (1 command, 1 agent)
# → mcp.json: 0 servers reachable, 0 errors

# Validate the plugin manifest against the Claude Code schema
npx harness validate --plugin
# → plugin.json schema: ok
# → commands/plan.md: frontmatter ok
# → agents/default.md: frontmatter ok

# Load the workspace as a plugin in a one-off Claude Code session
claude -p --plugin-dir . "summarize the repo and propose a first task"

You can also pass extra flags through to the underlying metaharness call — anything after the harness name is forwarded:

npx @metaharness/claude-code@latest my-bot --skip-install --quiet

FAQ

Q: Is this different from running claude init? A: Yes. claude init writes a single CLAUDE.md into an existing repo. This scaffolds a fresh, opinionated workspace: settings, commands, agents, MCP config, and a loadable plugin manifest, all wired together and validated by harness doctor.

Q: Do I need an Anthropic API key to scaffold? A: No. Scaffolding is offline. You only need a key when you actually run claude against the workspace.

Q: Can I use this as a Claude Code plugin instead of a workspace? A: Yes — that's why plugin.json is at the root. Run claude -p --plugin-dir <path-to-folder> and the commands and agents register into your session.

License

MIT. Built on metaharness (https://www.npmjs.com/package/metaharness).

Deep-dive

Full explainer gist: https://gist.github.com/ruvnet/79808ffb89eeb080b1c00bf79fba33e1