Questions, answered
How do I install AI tools across different editors and AI clients?
Run the metahub installer once, then install anything with a single command: mh install <slug>. The installer also connects the metahub MCP server to every editor it detects, so you can skip the terminal entirely and ask your assistant in plain language instead. The same artifacts work across Claude Code, Cursor, Claude Desktop, VS Code, Zed, Windsurf, Continue, Cline, Goose, Codex CLI, and Antigravity, so you do not need a different install path per editor.
What is metahub?
metahub is an open registry for AI tools: skills, MCP servers, agents, and plugins that install into AI coding editors with one command. Every listed artifact is pinned to a specific commit and carries a public evaluation report, and publishers get a free dashboard showing how their tool behaves in the wild. It is free to browse, install, and publish.
How does metahub check that a tool is safe and works correctly?
Every artifact runs through Assay, our open-source evaluation framework, before it is listed. Assay executes the tool against behavioral scenarios in a sandbox and scores it on correctness, instruction adherence, safety, and latency, alongside static checks on the repository such as license and documentation. Results are public on each listing, including the checks that failed, so you can judge a tool on evidence rather than on its README.
Where can I find AI skills and plugins that have already been evaluated?
Browse or search the registry at registry.metahub.ai. Every artifact in the catalog has been evaluated before listing, and each listing links its full public report. You can filter by kind (skill, MCP server, agent, plugin), by category, and by whether the publisher is a verified owner.
Can I publish my own AI skill or agent, and what do I get?
Yes, and it is free. Sign in with GitHub at developer.metahub.ai, point metahub at a public repository, and it detects the artifacts inside, evaluates them, and publishes them. In return you get a public evaluation report, a listing people can install in one command, and a live dashboard of real usage. Publishing does not transfer any rights: your code stays in your repository under your license.
What usage data does metahub give publishers?
Publishers see invocations over time, p50 and p95 latency, error rate, which models and tools the artifact was used with, and recent errors. The data comes from an opt-out telemetry SDK that reports anonymous usage. It never sends prompts, file contents, or any user data, and end users can switch it off.
Is Assay open source, and can I run it in my own CI?
Yes. Assay is released under Apache-2.0 and published on npm, so you can run the same evaluation locally or in your CI pipeline that metahub runs before listing. Because the checks and scoring are open, you can see exactly why an artifact scored what it did, and open a pull request when you think a check is wrong.
How is metahub different from a marketplace or an awesome list?
A directory tells you a tool exists. metahub tells you whether it works: every artifact is evaluated before listing, pinned to the exact commit that was evaluated, and installable in one command across every supported editor. Listings carry public evidence rather than stars alone, and publishers get observability rather than just a link.
Does metahub check for prompt injection or destructive behaviour?
The evaluation includes a safety dimension: artifacts are exercised against adversarial scenarios and scored on how they respond, and unsafe behaviour lowers the score and is flagged on the public report. It is a meaningful signal, not a security guarantee. Because installs are commit-pinned and the source repository is linked from every listing, you can always read exactly what you are about to run.
What does commit-pinned mean, and why does it matter?
Every artifact is published at a specific commit SHA, and that is the exact revision you install. It matters because it makes an install reproducible and honest: you get the same code that was evaluated, not whatever the default branch happens to contain today. When a publisher ships a new version, it is evaluated again and the listing moves forward deliberately.