A docs-only MCP tool allowlist in Python, using Arcmira API documentation

[Zeal Caiden (zealous1)]

We maintain Arcmira, a YouTube transcript search API. Here is a small Python example for giving a Semantic Kernel agent access to our API documentation over Streamable HTTP, without an Arcmira account or a model call.

The useful detail is the tool allowlist: the documentation server also exposes a feedback submission tool. Instead of adding the entire MCP connector to the kernel, create a KernelPlugin containing only documentation search and documentation file reads. Keep the MCP context open for the lifetime of the agent run.

Install the versions we tested in a fresh environment:

python -m venv .venv
source .venv/bin/activate
python -m pip install 'semantic-kernel[mcp]==1.44.1' 'mcp==1.30.0'

Save this as docs_search.py and run python docs_search.py:

import asyncio
from semantic_kernel import Kernel
from semantic_kernel.connectors.mcp import MCPStreamableHttpPlugin
from semantic_kernel.functions import KernelPlugin


async def main():
    async with MCPStreamableHttpPlugin(
        name="arcmira_docs",
        url="https://arcmira.com/docs/mcp",
        load_prompts=False,
        request_timeout=30,
    ) as docs:
        plugin = KernelPlugin(
            name="arcmira_docs",
            functions=[
                docs.search_arcmira_api,
                docs.query_docs_filesystem_arcmira_api,
            ],
        )
        kernel = Kernel()
        kernel.add_plugin(plugin)
        result = await kernel.invoke(
            plugin_name="arcmira_docs",
            function_name="search_arcmira_api",
            query="How do I authenticate and search YouTube transcripts?",
        )
        print(result)


if __name__ == "__main__":
    asyncio.run(main())

We verified tool discovery, documentation search and a documentation file read with these versions. Exactly two tools were exposed to the kernel. The filesystem tool reads the documentation service, not local files. Semantic Kernel marks this connector experimental, and this dependency stack includes Azure prerelease packages. Model inference was not part of this check.

Full Semantic Kernel recipe · API documentation

There is also a Microsoft Agent Framework recipe, tested with agent-framework-core==1.19.0, mcp==1.30.0 and httpx==0.28.1. That version needs MCP 1.x, so keep it separate from frameworks requiring MCP 2.x.

The docs endpoint above does not search videos. Authenticated YouTube transcript research uses the separate Arcmira MCP server at https://mcp.arcmira.com/mcp; that is not what this example validates.

Feedback on the tool filtering and connection lifetime pattern is welcome. We maintain the recipes in our documentation so we can keep their tested versions and setup details current.