How we check

Launch7 doesn’t run, test or certify skills. It reads each skill’s source, writes down what it found in plain words, and links to the original so you can look for yourself.

What the check reads

For a skill, Launch7 reads every text file in its folder on GitHub — the SKILL.md instructions your AI loads, and its scripts, references and settings — at one recorded version (a commit). The current results were read on 5 October 2026. From that it notes:

• Instructions only, or scripts too

  Whether the folder holds programs your AI could run, and how many.

• Terminal commands and installs

  Whether the instructions tell your AI to run commands or install other software.

• Keys and outside addresses

  Whether it asks for an API key or token, and which outside web addresses it mentions.

• What it lets your AI do without asking

  Terminal access granted in advance, commands that run as soon as the skill loads, and hooks that keep running afterwards.

• Patterns worth a second look

  More than 40 of them: hidden or encoded text, prompt-injection phrases, downloading and running code in one step, reading passwords or keys and sending them out, switching off safety prompts, and commands that delete or persist.

• Whether it is still there

  If the repository has moved, been archived or gone, and when it last changed.

Each finding is graded on its own — worth knowing, worth reviewing before you install, or a potential risk — and shown with the file and line it came from, linked to that exact version on GitHub. Findings are never added up into a score. Files that are not text, or are very large, are listed by name only, and the skill is marked “Partly checked”. Skills added since the last check say “Not scanned yet”.

The install steps Launch7 shows fetch the exact version that was read, not whatever is on GitHub today.

For a connector (MCP server), Launch7 reads its entry in the official MCP registry: whether it runs on the publisher’s servers or on your computer, and which keys it needs.

What it does not do

• It does not run anything

  No skill or connector is installed or executed by Launch7. “Works with” follows each AI’s own documentation, not a test.

• It reads, it does not understand

  It matches patterns. Instructions reworded to avoid them, or code assembled while it runs, can slip through. It can also flag things that are fine — security tools often quote attacks — so each finding shows the line, for you to judge.

• A result describes one version

  The author can change the files after the check. Launch7’s install steps fetch the version that was read; copying the repository yourself gets whatever is there now.

• It is not a security audit

  “No known risk pattern found” is a statement about the check, not a guarantee about the skill.

The marks

Every list and every skill page uses the same few marks. There is no score and no green tick.

• You are here, or act here. The selected item, the primary button. The logo’s square.
• Works in this AI, according to its maker’s documentation.
• Partly, or only on some plans.
• Not supported — or a plain fact the check noted, with nothing to review.
• Worth reviewing before you install — scripts, terminal commands, API keys, outside addresses, or a pattern worth a look.
• Potential risk found — always shown with the file and line.
• Not checked yet. Unknown stays unknown.

Who published it

“From Anthropic”, “From Microsoft” and the like mean the skill comes from that company’s own GitHub account. That tells you who made it — not that it is safe, and not that the company reviewed this listing. Everything else is community-made and says which account it comes from. For connectors, the MCP registry confirms who controls the name; it does not review what the server does.

Something wrong?

If a listing is out of date, misleading or does something it shouldn’t, tell us and we’ll look: hello@launch7.in. Or start from the job: browse skills by job.