Try it today
Install the MCPX Skill
MCPX isn't part of the official MCP spec yet. In the meantime, you can use it right now by installing it as a skill in your AI coding agent. Once installed, your agent can discover and call any REST API that publishes an MCPX manifest — no MCP server required.
Installation
Two ways to install
Via the Skills CLI
If you use the skills.sh ecosystem, install with a single command. This works across Claude Code, Cursor, Codex, and 30+ other agents.
Ask your agent
Paste this into any AI coding agent — it'll fetch the skill and install it in the right place automatically.
How it works
What the skill does
Once installed, the skill teaches your AI agent a simple three-step workflow for connecting to any REST API with an MCPX manifest.
Discover
The agent fetches the MCPX registry to find available services. If a service isn't in the registry, the agent checks for a manifest at /.well-known/mcp.yaml on the service's domain.
Authenticate
The manifest declares how authentication works — API key or OAuth 2.0 with PKCE. The agent checks for stored credentials first. If none exist, it asks the user for an API key or walks them through the OAuth flow, then stores the credentials for future sessions.
Invoke
The agent reads the OpenAPI spec to understand which endpoints are available, what parameters they accept, and what they return. It calls standard REST endpoints with the user's credentials — no MCP server, no special transport.
Credentials
How credentials are managed
The skill instructs the agent to store API keys and OAuth tokens locally so you don't have to re-authenticate every session.
| Auth type | What happens |
|---|---|
| API key | Agent asks for the key once, stores it in ~/.claude/credentials/ |
| OAuth 2.0 | Agent directs user to the authorization URL, exchanges the code for tokens, and stores both access and refresh tokens |
The agent always confirms before calling endpoints that create, modify, or delete data. Read-only operations are preferred when just gathering information.