arcmira-docs:
https://arcmira.com/docs/mcp
Connect from an agent framework
Each recipe connects over Streamable HTTP and either calls docs search once or builds the toolset for your agent. None makes a model call or needs an Arcmira key. Each exposes only two tools,search_arcmira_api and query_docs_filesystem_arcmira_api, and leaves out the tool that submits documentation feedback. The filesystem tool reads the documentation service, not your local files. Keep the client or context open until the agent finishes, streamed calls included.
Every recipe passed live tool discovery and documentation search with the versions it names. Model inference was not tested. Running an agent incurs your model provider’s usual charges, and docs access does not grant video research. For YouTube transcript search from a framework, see CLI and SDKs.
- Vercel AI SDK
- Mastra
- OpenAI Agents SDK
- CrewAI
- LlamaIndex
- Haystack
- Microsoft Agent Framework
- Semantic Kernel
- LangChain JavaScript
- LangChain Python
- Pydantic AI
- Google ADK
Tested with Pass
@ai-sdk/mcp 2.0.60 and zod 4.6.5. Install those packages in your AI SDK project.import { createMCPClient } from '@ai-sdk/mcp';
import { z } from 'zod';
const docsClient = await createMCPClient({
transport: { type: 'http', url: 'https://arcmira.com/docs/mcp' },
});
try {
const tools = await docsClient.tools({
schemas: {
search_arcmira_api: { inputSchema: z.object({ query: z.string() }) },
query_docs_filesystem_arcmira_api: {
inputSchema: z.object({ command: z.string() }),
},
},
});
const result = await tools.search_arcmira_api.execute(
{ query: 'How do I authenticate and search YouTube transcripts?' },
{ messages: [], toolCallId: 'arcmira-docs-search' },
);
if (result.isError) throw new Error(JSON.stringify(result.content));
console.log(JSON.stringify(result));
} finally {
await docsClient.close();
}
tools to your generateText call inside the try block. The explicit schemas expose only the two read tools. See the AI SDK MCP guide for model setup and streaming cleanup.Tested with Pass
@mastra/mcp 2.1.0, @mastra/core 1.71.0 and zod 4.6.5. Install those packages in your Mastra project.import { MCPClient } from '@mastra/mcp';
import { noopObserve } from '@mastra/core/tools';
const docsClient = new MCPClient({
id: 'arcmira-docs',
servers: {
arcmiraDocs: { url: new URL('https://arcmira.com/docs/mcp') },
},
});
try {
const discovered = await docsClient.listTools();
const tools = Object.fromEntries(
Object.entries(discovered).filter(([name]) =>
['arcmiraDocs_search_arcmira_api',
'arcmiraDocs_query_docs_filesystem_arcmira_api'].includes(name),
),
);
const search = tools.arcmiraDocs_search_arcmira_api;
if (!search?.execute) throw new Error('Arcmira docs search is unavailable');
const result = await search.execute(
{ query: 'How do I authenticate and search YouTube transcripts?' },
{ observe: noopObserve },
);
console.log(JSON.stringify(result));
} finally {
await docsClient.disconnect();
}
tools to your Mastra Agent inside the try block. Mastra prefixes tool names with the server name you configure. noopObserve supplies the execution context for this direct call without a tracing span. See the Mastra MCP guide for agent setup.Install Pass
openai-agents==0.22.3 and mcp==2.2.0, the versions this was tested with.import asyncio
from agents.mcp import MCPServerStreamableHttp, create_static_tool_filter
async def main():
async with MCPServerStreamableHttp(
name="arcmira-docs",
params={"url": "https://arcmira.com/docs/mcp"},
client_session_timeout_seconds=30,
tool_filter=create_static_tool_filter(allowed_tool_names=[
"search_arcmira_api", "query_docs_filesystem_arcmira_api",
]),
) as docs:
await docs.list_tools()
result = await docs.call_tool("search_arcmira_api", {
"query": "How do I authenticate and search YouTube transcripts?",
})
if result.is_error:
raise RuntimeError(result.model_dump_json())
print(result.model_dump_json())
asyncio.run(main())
mcp_servers=[docs] to your Agent inside the async with block. See OpenAI’s integrations guide for agent setup and tracing controls.Tested with Pass
crewai and crewai-tools 1.15.22. Install crewai-tools[mcp] in your CrewAI project. This release requires MCP 1.x; let its dependency constraint pick the version.from crewai_tools import MCPServerAdapter
with MCPServerAdapter(
{"url": "https://arcmira.com/docs/mcp", "transport": "streamable-http"},
"search_arcmira_api",
"query_docs_filesystem_arcmira_api",
) as tools:
result = tools["search_arcmira_api"].run(
query="How do I authenticate and search YouTube transcripts?"
)
print(result)
tools to your CrewAI Agent inside the with block. See the CrewAI Streamable HTTP guide for agent setup.Install Pass
llama-index-tools-mcp==0.6.0 and llama-index-core==0.14.25, the versions this was tested with. The MCP package uses the MCP 2.x client.import asyncio
from llama_index.tools.mcp import McpToolSpec
from mcp import ClientSession
from mcp.client.streamable_http import streamable_http_client
async def main():
async with streamable_http_client("https://arcmira.com/docs/mcp") as (read, write):
async with ClientSession(read, write) as session:
await session.initialize()
tools = await McpToolSpec(
client=session,
allowed_tools=["search_arcmira_api", "query_docs_filesystem_arcmira_api"],
).to_tool_list_async()
search = next(t for t in tools if t.metadata.name == "search_arcmira_api")
result = await search.acall(
query="How do I authenticate and search YouTube transcripts?"
)
if result.raw_output.is_error:
raise RuntimeError(str(result))
print(result)
if __name__ == "__main__":
asyncio.run(main())
tools to your LlamaIndex agent while both contexts are open. See the LlamaIndex MCP guide for framework setup.Install Pass
mcp-haystack==1.5.1 and haystack-ai==3.2.0.from haystack_integrations.tools.mcp import MCPToolset, StreamableHttpServerInfo
def main():
docs = MCPToolset(
server_info=StreamableHttpServerInfo(
url="https://arcmira.com/docs/mcp", max_retries=0
),
tool_names=["search_arcmira_api", "query_docs_filesystem_arcmira_api"],
)
try:
docs.warm_up()
search = next(tool for tool in docs if tool.name == "search_arcmira_api")
print(search.invoke(query="How do I authenticate and search YouTube transcripts?"))
finally:
docs.close()
if __name__ == "__main__":
main()
tools=docs to your Haystack agent and keep the toolset open until it finishes. See the Haystack MCP reference for framework setup.Install Pass
agent-framework-core==1.19.0, mcp==1.30.0 and httpx==0.28.1. This framework version expects MCP 1.x.import asyncio
from agent_framework import MCPStreamableHTTPTool
async def main():
async with MCPStreamableHTTPTool(
name="arcmira_docs",
url="https://arcmira.com/docs/mcp",
allowed_tools=["search_arcmira_api", "query_docs_filesystem_arcmira_api"],
load_prompts=False,
request_timeout=30,
) as docs:
result = await docs.call_tool(
"search_arcmira_api",
query="How do I authenticate and search YouTube transcripts?",
)
print("\n".join(item.text for item in result if item.text))
if __name__ == "__main__":
asyncio.run(main())
tools=docs to your agent run while the context is open. See the Microsoft Agent Framework MCP guide for agent setup.Install Use this kernel with your Semantic Kernel agent while the MCP context is open. Add only the filtered
semantic-kernel[mcp]==1.44.1 and mcp==1.30.0, in a separate environment from frameworks that need MCP 2.x. Semantic Kernel marks this connector experimental, and its dependencies include Azure prerelease packages.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())
plugin; adding docs directly also exposes the documentation feedback tool. See the Semantic Kernel MCP guide for agent setup.Tested on Node.js 22.22 with the 1.x adapter API.Save this as Pass the filtered
npm install @langchain/mcp-adapters@1.1.4 @langchain/core@1.2.12 @langchain/langgraph@1.4.18
docs-search.mjs, then run node docs-search.mjs.import { MultiServerMCPClient } from "@langchain/mcp-adapters";
const client = new MultiServerMCPClient({
arcmiraDocs: {
transport: "http",
url: "https://arcmira.com/docs/mcp",
},
});
try {
const allowed = new Set([
"search_arcmira_api",
"query_docs_filesystem_arcmira_api",
]);
const tools = (await client.getTools()).filter(tool => allowed.has(tool.name));
const search = tools.find(tool => tool.name === "search_arcmira_api");
if (!search) throw new Error("Arcmira documentation search is unavailable");
const result = await search.invoke({
query: "How do I authenticate and search YouTube transcripts?",
});
console.log(JSON.stringify(result, null, 2));
} finally {
await client.close();
}
tools to your LangChain or LangGraph agent inside the try block. Version 2.x replaces MultiServerMCPClient with MCPAdapter, and this example has not been tested with it. See the 1.x MCP guide and the 2.x migration guide.Install Pass
langchain[mcp]==1.4.2, the version this was tested with. The langchain.mcp API is in beta.import asyncio
from langchain.mcp import MCPAdapter
async def main():
read_tools = {"search_arcmira_api", "query_docs_filesystem_arcmira_api"}
async with MCPAdapter("https://arcmira.com/docs/mcp") as adapter:
tools = [tool for tool in await adapter.list_tools() if tool.name in read_tools]
search = next(tool for tool in tools if tool.name == "search_arcmira_api")
result = await search.ainvoke({
"type": "tool_call",
"id": "arcmira-docs-search",
"name": search.name,
"args": {"query": "How do I authenticate and search YouTube transcripts?"},
})
if result.status == "error":
raise RuntimeError(result.content)
print(result.content)
asyncio.run(main())
tools to create_agent with your model configuration inside the async with block. See the LangChain MCP guide for agent setup.Tested with Use
pydantic-ai-slim[mcp] 2.51.0.from pydantic_ai.mcp import MCPToolset
read_tools = {"search_arcmira_api", "query_docs_filesystem_arcmira_api"}
docs_tools = MCPToolset("https://arcmira.com/docs/mcp").filtered(
lambda ctx, tool: tool.name in read_tools
)
toolsets=[docs_tools] when you construct your Agent. See the Pydantic AI MCP client guide for connection lifecycle and model setup.Tested with Use
google-adk 2.10.0.from google.adk.tools.mcp_tool import McpToolset
from google.adk.tools.mcp_tool.mcp_session_manager import StreamableHTTPConnectionParams
docs_tools = McpToolset(
connection_params=StreamableHTTPConnectionParams(
url="https://arcmira.com/docs/mcp"
),
tool_filter=["search_arcmira_api", "query_docs_filesystem_arcmira_api"],
)
tools=[docs_tools] when you construct your LlmAgent. Standalone callers must run await docs_tools.close() in a finally block. See the Google ADK MCP guide for agent setup and connection cleanup.