Choose the target agent before adding a skill
Running npx skills add vercel-labs/agent-skills asks you which instruction packages to add and which coding agents should receive them. That second choice explains why Vercel Labs' Skills CLI exists. A team using Claude Code alongside Cursor or OpenCode needs those instructions in the locations each tool recognizes. Skills handles discovery and placement, while the agent still decides how to use the material.
A skill is a folder of instructions, sometimes with scripts or reference files. Installing one changes what an agent can read and potentially execute. I would use Skills to keep a reviewed collection available across a mixed toolchain. Someone who only needs a personal instruction file for a single agent has less reason to add another manager. The CLI's convenience becomes useful when placement itself is repetitive.
What happened when we ran it
Our sandbox installed commit 13e4063 in 8 seconds, built it in 15 seconds, and completed the Vitest step in 75 seconds: 964 passed and 0 failed out of 964. The October 10, 2026 run recorded 126 dependencies and 94 MB on disk. Those are encouraging results for someone considering work on the CLI's source.
The checkout contained 138 files, roughly 32,542 source lines, and occupied 1.2 MB. The recorded configuration was 3 CPUs, 8 GB of RAM, and image lab-node:22. That image name does not verify the runtime version. Our signals recorded 3 CI workflow files, a tests directory, and no Dockerfile. None of these observations establishes that an installed instruction package improves an agent's answers.
The test result belongs to that commit and that sandbox. It does not demonstrate compatibility with every supported agent, validate every public skill, or settle the current lockfile reports. Treat it as evidence that the measured source workflow completed successfully, then check the actual handoff to your chosen agent separately.
Project scope gives you a smaller change to inspect
The README makes project installation the default, with global scope available through -g. The command reference also lets you name a particular skill and target agent. That is the route I would start with: add a known package to a disposable project, inspect the resulting files, and confirm that the agent discovers it before expanding the collection.
Interactive installation offers symlinks to a canonical copy or independent copies for each agent. Symlinks keep one source file behind several agent paths. Copies are useful where links are unsupported, but give you more files to track. Choose deliberately if teammates move between operating systems or use tools that rewrite instructions. A successful CLI exit is only part of a useful handoff.
The current README requires Node.js 22.20 or later. Public sources do not require a paid model-service credential simply to retrieve instructions. Private repositories use the authentication configured for their URL, such as Git credentials or SSH. Those access requirements belong to the source repository, while the selected coding agent has its own account and runtime requirements.
Lockfile restoration leaves agent links to you
The command reference says experimental_install restores supported lockfile sources into .agents/skills/. It also says that command does not create agent-specific links. A developer expecting a fresh checkout plus a restore command to reproduce every teammate's agent directories should stop here and design that missing step explicitly.
There is a broader concern in open issue 2560. The report describes update announcing that a skill is current even after its directory has been deleted, because the decision compares hashes without checking disk. It also documents a corrupt lock being treated as empty. These are upstream reports about current behavior, separate from our earlier sandbox result. The issue's proposed changes are not shipped guarantees.
For a shared repository, review both the lockfile and the installed directory after changes. Keep a recoverable copy before bulk operations. This is a concrete reason to postpone adopting Skills as an unattended restoration mechanism, even if you are happy to use its interactive add command. Experimental npm synchronization also remains explicitly subject to changing behavior.
OpenSkills and Anthropic's collection answer narrower needs
OpenSkills writes an available-skills listing into AGENTS.md. Choose it when that file is already your agent's entry point and you want an explicit read command for loading instructions. This is a different integration choice, rather than evidence that either project's skill content produces better work. Our sandbox supplied no comparison of agent output.
Anthropic's Skills repository can be registered as a Claude Code plugin marketplace. Choose that direct route when its example or document skills meet your needs and Claude Code is your only target. Read the individual licenses: Anthropic distinguishes its open-source examples from source-available document skills. Neither collection licensing nor instruction quality transfers automatically to other repositories found through a directory.
Maintenance is active, and restore behavior needs scrutiny
The last push was October 19, 2026, and issue 2560 had activity on October 22. Together, those dates show ongoing work while exposing an unresolved operational boundary. They support trying the CLI with supervision. They do not justify trusting lockfile restoration without inspecting the resulting directories.
Telemetry is enabled by default, and the README documents DISABLE_TELEMETRY=1 or DO_NOT_TRACK=1 to disable telemetry and security-audit requests. Set your policy before adding skills. Then choose one instruction package you can read completely, install it for the intended agent, and verify the files that agent will actually load.