With most agent frameworks, tool calls run in a loop like this:
agent → tool → agent → tool → agent → tool
Each step goes back through the model. For tools returning large payloads, this adds latency and burns tokens.
So I built a small tool that converts an MCP server into a Python skill.
It introspects the server and generates a typed Python class where each tool becomes an async method.
Example:
issue = await github.get_issue(issue_id)
docs = await search.web_search(issue.title)
This also enables progressive disclosure. Tools can return large responses, but the agent can filter or aggregate data in code instead of sending everything back into the model context.
The generated skills include a `SKILL.md` file and can be reused later, so useful workflows can be saved and run again as scripts.
Still a work in progress, but already useful in my own setup.
Would love feedback from others experimenting with MCP or similar patterns.