Cusp Hub

The Hub is where you'll find and share things that extend Cusp: skills, agents, MCP servers, hooks, commands, workflows, output styles, language servers and themes, plus blueprints that set up a whole project in one step. It's built around three questions you should be able to answer before you install anything: what will this be allowed to do, how well does it work with my model, and is this exactly the version I reviewed?

Extensions stay compatibleCopy Anchor Link

A Hub extension is an ordinary plugin folder (.claude-plugin/plugin.json), an MCP server.json, a SKILL.md or an agent file. Anything Cusp adds lives in an optional sidecar, .cusp-plugin/hub.json, which other tools ignore. The same folder works as a plugin in other agents that read this format and as a Hub extension. Gemini CLI extensions and .mcpb bundles are read as well.

Capability labelsCopy Anchor Link

Every extension carries a capability manifest, shown before you install it like a nutrition label:

Cusp also infers capabilities from the extension's contents and flags anything the publisher didn't declare. Once installed, the label isn't just information: it's enforced by the same permission engine and sandbox that govern everything else. Secrets go in the OS keychain, never into settings files.

Per-model scoresCopy Anchor Link

An MCP server or skill that works well with one model can confuse another. Hub extensions can ship an eval suite so the Hub can report results per model family (Claude, GPT, Gemini, GLM, Qwen, Kimi, DeepSeek, MiniMax, Mistral, Grok, Llama). Suites aren't run yet, so no scores are published. Publishers can also ship tool descriptions tuned per family; Cusp doesn't apply them yet.

BlueprintsCopy Anchor Link

A blueprint sets up a project in one step. It bundles:

• Extensions to install.
• Integrations to connect, such as GitHub or Linear.
• Crew seats to hire, with their roles and schedules.
• Workflows.
• Project settings: model, output style, permission rules and non-secret environment variables.
• A welcome message and a suggested first prompt.

Applying a blueprint shows everything it will add and every permission rule it will set before anything changes.

Pinned and verifiable installsCopy Anchor Link

Installs are recorded in a lockfile with integrity hashes, so the same project installs the same bytes on every machine. Tool definitions are pinned too: when an extension ships a tool snapshot and its MCP server later changes a tool's description, Cusp hides that tool until you approve the change instead of silently trusting the new text. The registry signs every version it publishes, and extensions are scanned for risky patterns before install.

Plugins todayCopy Anchor Link

You can install plugins and marketplaces in the .claude-plugin format with the CLI now:

cusp plugin marketplace add owner/repo     # or a git URL, a URL, or ./dir
cusp plugin list --available
cusp plugin install some-plugin@some-marketplace
cusp plugin install ./my-plugin --project  # this project only
cusp plugin disable some-plugin
cusp plugin validate ./my-plugin

Plugins can add commands, agents, skills, hooks, MCP and language servers, and output styles. They can run code on your machine, so install only plugins you trust. These plugins don't carry Hub capability labels or scores yet.

NextCopy Anchor Link

• Skills, Agents, Hooks and MCP: what an extension can contain.
• Publish to the Hub: prepare a package, its capability label and its tool snapshot.
• Security model: how capabilities are enforced.