Scenario: multiple harnesses

Several AI coding tools working on the same repository. This page is the multi-harness topic — per-harness getting started lives in the agent harness guide.

What you do

Each person installs the plugin in their own harness exactly as the harness guide describes. Claude Code, Codex, Copilot, and Cursor each resolve the same canonical skill bytes, so two tools never see different guidance.
Pi (and any npm-consuming host) installs the published package with an exact pin recorded in .pi/settings.json:
Terminal window
```
pi install -l npm:@cratis/ai-fundamentals@0.10.0
```
The repository commits .cratis/ai.json as the one neutral record of intended scope and version — harness-neutral by construction, because it names profiles (cratis/chronicle/kotlin), not host-specific things. The harnesses array inside it lists which hosts this repository expects the skills to reach.
Per-host locks stay in their own settings files (.pi/settings.json, your host’s plugin list). No host reads another host’s lock.

What you get

One guidance corpus across every tool. All marketplaces resolve the same skills/ directory on the Cratis/AI default branch; adapters change packaging and discovery, never skill content.
One reviewed scope. A pull request changes .cratis/ai.json once, and every harness’s humans see the same diff.
No cross-contamination. Skills are passive; no harness writes into another harness’s configuration.

When you don’t need this

If everyone on the repository uses the same single harness, this page adds nothing — the team repository contract alone is enough.

Status

Per-host package availability differs today: marketplace hosts follow Cratis/AI main; npm/Pi carries only the unsupported @cratis/ai-fundamentals 0.x evaluation; Kiro, Junie, and Gemini CLI are pending (see the harness guide). The versioned flow that would pin every harness to one reviewed release is designed but not published yet.