AI tools are exploding.
The infrastructure isn't.
Thousands of skills, MCP servers, agents, and plugins, scattered across GitHub with no install story, no quality bar, and no way to tell what actually works. metahub is building the missing infrastructure, in the open. We're starting with two core pieces: a universal registry that installs anywhere, and the open-source eval framework that powers it.
One command, either way. Then mh install <anything> works in any harness: Claude Code, Cursor, Codex CLI, Gemini CLI, and 9 more.
What it does: the shell script (a plain file you can read before running) installs the mh CLI and MCP server via your package manager and adds them to PATH; npx runs the same setup with nothing installed globally. Both connect the MCP server to the harnesses they detect and tell them to look on metahub first (skippable, reversible). Installed tools report anonymous, opt-out usage stats to their publisher. Never your prompts, never your files. Full details.
The registry
A registry for AI tools. Every artifact is pinned to an exact commit, scanned and evaluated before it goes public, and installs with one command into any harness, 12 and counting. Browse by use case, check the eval report, read real reviews, install. Free.
-
01Find a tool at
registry.metahub.ai, by category, search, or the official shelf. -
02Run
mh install <slug>. Commit-pinned: you get exactly what was evaluated. - 03It works across your editors. Review it so the next person knows.
Tested, then listed
You cannot tell whether an AI tool works by reading it. Stars measure attention, not behaviour, and a README is a claim its author is grading. So metahub runs the thing, with Assay, our open-source eval framework: every artifact is executed in a sandbox against real prompts and graded on the transcript, then probed with prompt injection, out-of-scope bait and destructive requests. The score, the failures and the commit it was measured at are published on the listing, whether or not they flatter us.
- ScoredCorrectness, instruction-following, safety and latency, each out of 10, from a real run rather than a checklist someone ticked.
- ProbedPrompt injection, out-of-scope bait, destructive requests. A tool that complies loses points for it.
- PinnedEvery score names the commit it was measured at, so what you install is the thing that was graded.
- OpenThe rubric, the probes and the judge prompts are public as Assay, Apache-2.0. Run it on your own repo before you publish, or send a PR when a rubric is wrong.
See it in the wild
Publishing an AI tool usually means never hearing from it again. You get stars, maybe an issue. You never learn whether it actually fired, how slow it was, or which model it ran under. Every artifact published to metahub gets a free, live dashboard from its first real invocation. Why? Because you can't improve what you can't measure.
- 01Wrap the artifact with the open SDK. One line, no config, no separate account.
- 02Publish the repo. The dashboard starts filling from the first real invocation, not from a demo.
- 03Read invocations, p95 latency, model mix, handoffs and error rate, plus reviews that only verified installs can leave.
Built in the open
The registry is the first piece, not the product. The plan is an ecosystem of tools that make AI extensions trustworthy: evaluated on the way in, observable in the wild, and easy to build.
The registry
Catalog, evals, reviews, one-command install into any harness.
The eval framework
The thinking behind every listing, released as Assay: read the checks, run them on any repo, and send a PR when a rubric is wrong.
Publisher observability
Live invocations, latency, model mix and error rate. Free, and shipped.
The wider ecosystem
More tools are already in build on top of the registry. Each one gets announced when you can actually use it, not before.
Get involved
While you browse
Every page of the registry has a Feedback button sitting in the corner. Two clicks, no account needed, and the page you were on comes attached. Bug, idea, or just a feeling: we read every single one.
Break it for us
You're early, which means your half-formed complaint is worth more than a thousand stars. Taking 20 minutes to kick the tires is plenty.
-
Browse
registry.metahub.ai. Search something you'd actually use. -
Install one artifact with
mh install, or straight from your AI client. - Leave a review on it. Sign in with GitHub if you want the verified badge.
- Anything confusing, broken, or slow: hit the Feedback button, bottom-right of every page. It lands on our desk with the page attached.
Build with us
Publishers and contributors get the ecosystem's best seats: distribution today, a say in what gets built next.
-
Publish your skill, MCP server, or plugin at
developer.metahub.ai. GitHub repo in, live listing out, in about 90 seconds. - Get free observability the day you publish: invocations, latency, reviews.
- Watch the eval framework land in the open. Early testers hear about it first.
Common questions
The four we get asked before anyone runs the install command. What metahub is, what it makes, and what it costs are answered on the full FAQ; installing from the catalog and publishing into it are answered in the registry FAQ.
How do I install metahub?
One command, and a plain shell script you can read before you run it:
curl -fsSL https://metahub.ai/install.sh | sh
Or, with nothing installed globally:
npx @metahub-ai/mh bootstrap
The script installs the mh CLI and the metahub MCP server through your package manager and adds them to PATH; npx runs the same setup on the spot. Both connect the MCP server to the AI harnesses they find on your machine and add a short note to each one's instructions so it looks on metahub first, a step you can skip and reverse later. After that a single mh install command works in any supported harness.
Which AI clients does metahub work with?
Thirteen, including Claude Code, Cursor, Codex CLI, Gemini CLI, Antigravity, and opencode. The same install command wires an artifact into whichever of them you use, so a skill, MCP server, agent, or plugin does not need separate setup instructions per harness.
What does metahub send back when I install a tool?
Anonymous, opt-out usage statistics, and nothing else. Installed tools report invocation counts, latency, and error rates back to the person who published them. Never your prompts, never your files. You can switch the telemetry off.
How does metahub decide what gets listed?
Every artifact is evaluated before it reaches the catalog. The evaluation runs on Assay, our open-source framework released under Apache-2.0, so both the checks and the scoring are public. You can run the same evaluation on your own repository, and open a pull request when you think a check is wrong.
Still stuck, or want to know who is building this? Get in touch or read what metahub is for.