AI ecosystem support
Cratis authors one corpus and exposes it through each harness’s native folders. The adapter changes how a harness discovers the content, not what the rule or skill says. What a harness receives depends on the channel you install through.
Support matrix
Section titled “Support matrix”| Harness | Managed install (cratis ai install) | Native channel | Native channel content |
|---|---|---|---|
| Claude Code | CLAUDE.md, .claude | cratis marketplace plugin | Skills |
| OpenAI Codex | AGENTS.md, .agents/skills | cratis marketplace plugin | Skills |
| GitHub Copilot | .github | cratis marketplace plugin | Skills |
| Cursor | .cursor | cratis marketplace plugin | Skills |
| OpenCode | AGENTS.md, .opencode | None | Use the managed install |
| Pi | AGENTS.md, .pi | @cratis/pi on npm | Skills, rules, prompts, agents, subagent tool, quality hooks |
The managed install gives each selected harness the integrations it supports (rules, skills, prompts, and agents; Codex gets only AGENTS.md and skills) from one committed copy, filtered by profile and language. It is the
recommended path for a repository that more than one tool or person works in.
The agent harness guide lists exactly what it creates for each
harness.
Harnesses not in this table are not supported by the CLI or the marketplace
manifests. A tool that reads AGENTS.md may still pick up the root
instructions, but Cratis does not test or document that.
Versions
Section titled “Versions”| Channel | Where it reads from | Can you pin it? |
|---|---|---|
| Managed install | Current main of Cratis/AI, or a local checkout with --source | Only by pointing --source at a checkout of a known commit |
| Marketplace plugins | Current default branch of Cratis/AI | No |
@cratis/pi | Published npm releases | Yes, by package version |
Cratis AI distribution explains the ownership and update rules behind these channels.
Quality
Section titled “Quality”The Cratis/AI repository checks corpus structure, profile composition, the
harness adapters, and the Pi package in CI, and loads the corpus through Pi’s
resource loader to prove the skills, prompts, and extensions are discovered.
Per-skill checks are static: they confirm that a skill contains required
guidance. They do not run a model against the skill, compile its examples, or
measure how well an agent performs with it.