The Claude Code plugin

The plugin puts three things into Claude Code with two commands: the Harness Lab server, a skill that reviews your own agent’s harness against the research, and a skill that compares harnesses on a topic. It is version 1.0.0. Claude Code 2.1.224 or later is needed.

1. 1. Make a key and keep it in the environment

   Make a key in Account, Connect your agent and keep it in HARNESSLAB_API_KEY, as in the quickstart. The plugin’s server reads the key from that variable, so the key is in no file. Start Claude Code from a terminal that has it.

2. 2. Add the marketplace

   A marketplace is a list of plugins. Claude Code can add one from a plain HTTPS address, so this site serves it and you need no Git account:

   In Claude Code

   /plugin marketplace add https://harnesslab.co/claude/marketplace.json

3. 3. Install the plugin

   In Claude Code

   /plugin install harness-lab@harnesslab

   Then run /reload-plugins (or start a new session) and type /mcp. The server is listed as plugin:harness-lab:harnesslab; ask for whoami to see what it connects as.

4. 4. Use it

   /harness-lab:harness-review

   Reviews your agent’s harness. It builds a short map of your harness files (the loop, tools, prompts, context and compaction, retries, stop conditions, sub-agents, memory, evals and model settings), shows it to you, and asks before it goes on. Then it asks the research tools about each area and reports findings as a table: the file and line in your code, the research topic, the evidence with its status, and a suggested change. It gives no single score.

   /harness-lab:compare

   Compares two to six harnesses on one topic and shows a table with the evidence links and the date of the knowledge base.

   Claude can also start either skill by itself when you ask for a harness review or a comparison.

What leaves your machine

The review skill reads your files with Claude Code’s own read tools, on your machine. What it sends to Harness Lab is only the short descriptions it writes itself (for example “compaction summarises at 80% of the window”) and topic and harness names. Its instructions forbid sending code, prompt text, whole files, .env files, keys or secrets, and it must show you the map before it asks anything. Claude Code’s own model sees your files, as it does in any session. Everything the tools return is research data, not instructions.

The files

• The marketplace: https://harnesslab.co/claude/marketplace.json. It lists one plugin, harness-lab, as a zip.
• The plugin: https://harnesslab.co/claude/harness-lab.zip. Claude Code checks it against the SHA-256 in the marketplace file, so a changed download is refused. Its SHA-256 is b2f65d7951269cc6cc7edcf156878fea3f5e8730ed6c89b1331347ba1d9ed206.
• The plugin’s server is https://app.harnesslab.co/mcp with the header Authorization: Bearer ${HARNESSLAB_API_KEY}.

A plugin is updated when its version changes. Run /plugin marketplace update harnesslab to fetch a new one.

Good to know

• If you also added the server with claude mcp add, Claude Code connects once: it matches a plugin and a server by address. See Troubleshooting.
• The plugin’s tools are named mcp__plugin_harness-lab_harnesslab__<tool> in permission rules.
• No key, or the variable not set, shows as a failed connection in /mcp; the skills tell you so instead of guessing.