Your AI assistant knows nothing about your private documentation. It calls a cloud server for every answer, and when that server is unavailable or you hit a rate limit, you get nothing. Skills packages fix that. A single file carries everything your AI needs to answer questions about your docs, and it runs entirely on your machine.
What Is Inside a Skills Package
A Skills package is one file. It contains your documentation content split into searchable segments, vector embeddings pre-computed for each segment, page titles and URLs, and a manifest with version and source metadata.
No model is required at query time. The embeddings are already computed. The AI plugin loads the file, runs a local similarity search, and returns results in under 100ms.
The format is open and documented, so you can inspect the contents or build tooling on top of it directly.
When Skills Packages Beat the MCP Server
Here is the honest answer: use the MCP server for daily development work. Use Skills packages for everything else. These are the four cases where a file beats a live server.
Offline and air-gapped environments. Development machines without stable internet, field deployments, or servers that cannot reach external services. The package works anywhere a file can be read.
Distribution alongside a product. Ship a CLI tool or desktop app? Include the package as a single file in your distribution. Your users get accurate documentation search without needing a Yavy account.
Sharing with contractors and external teams. Send the file, not your credentials. The recipient gets the knowledge with zero access to your account.
Reproducible snapshots for testing and CI/CD. Pin a specific documentation version as a build artifact. Regenerate it on each documentation release. That way, the package version and the documentation version are always in sync.
That last point matters for CI/CD specifically. Automated pipelines cannot authenticate interactively, and they need deterministic inputs. A file artifact is both.
Generating, Sizing, and Keeping Packages Current
From any Yavy project that has finished indexing, open project settings and click "Export Skills Package." You choose which sources to include, whether to bundle page content alongside the embeddings, and whether to use standard or compressed output.
Small projects finish in seconds. A 200-page documentation site takes under a minute and produces a file of 50 to 80 MB. Compressed format cuts that roughly in half, with a one-time decompression cost on first query of a second or two. If you are bundling a package inside a product you ship, compressed is worth it. For a developer's local machine, standard format avoids the loading delay.
The Yavy API exposes package generation as an endpoint, so you can wire it into a release pipeline. For teams, the right pattern is to regenerate on each documentation publish and attach the file as a versioned build artifact. Developers pull the latest package the same way they pull any other dependency.
A Skills package is a snapshot. It does not update itself. For individuals, regenerating every few weeks is usually enough. Yavy shows a timestamp on each package so you know exactly how stale the knowledge is.
Using a Package with Claude Code
The Yavy Claude Code plugin loads Skills packages with one config change. Point the --package argument at your file:
{
"mcpServers": {
"yavy-local": {
"command": "yavy-claude",
"args": ["--package", "./docs/my-project.skills"]
}
}
}The plugin exposes the same search_docs tool interface as the live MCP server. Claude sees no difference. Queries run locally, results come back instantly, and no data leaves your machine.
MCP Server vs. Skills Packages
| MCP Server | Skills Package | |
|---|---|---|
| Internet required | Yes | No |
| Always current | Yes | Requires regeneration |
| Query latency | 1-3 seconds | Under 100ms |
| Share without account access | No | Yes |
| Works in CI/CD | Yes (API) | Yes (file artifact) |
| Storage footprint | None (server-side) | 30-80 MB per project |
Neither is universally better. Most teams run the MCP server during development and include a Skills package in their release pipeline. If you have an existing Yavy project, exporting your first package takes about 30 seconds. Install the Claude Code plugin, point the config at the file, and you have local documentation search running in every Claude conversation, online or off.