Add it to your client config<\/strong><\/li>\n<\/ol>\n\n\n\nExample setup:<\/p>\n\n\n\n
mcpServers: {<\/p>\n\n\n\n
composio: {<\/p>\n\n\n\n
url: “https:\/\/mcp.composio.dev\/your-instance\/sse”,<\/p>\n\n\n\n
transport: “streamable_http”,<\/p>\n\n\n\n
auth: {<\/p>\n\n\n\n
apiKey: COMPOSIO_API_KEY,<\/p>\n\n\n\n
},<\/p>\n\n\n\n
}<\/p>\n\n\n\n
}<\/p>\n\n\n\n
Composio handles the hard stuff: auth, schema generation, scaling, and tool maintenance. You just connect and use.<\/p>\n\n\n\n
Option 2: Connect your own hosted MCP Servers<\/h3>\n\n\n\n
Already built your own server and deployed it somewhere? Great! You can connect it just like you did with Composio.<\/p>\n\n\n\n
Just add another entry to your config:<\/p>\n\n\n\n
mcpServers: {<\/p>\n\n\n\n
myTools: {<\/p>\n\n\n\n
url: “https:\/\/myhost.com\/mcp”,<\/p>\n\n\n\n
transport: “streamable_http”,<\/p>\n\n\n\n
}<\/p>\n\n\n\n
}<\/p>\n\n\n\n
No need to change your code or redeploy the client. LangGraph\u2019s MCP client handles it automatically.<\/p>\n\n\n\n
Here\u2019s the best part: you don\u2019t have to choose between local, hosted, or managed servers. You can use all of them at once<\/strong>.<\/p>\n\n\n\nExample combined config:<\/p>\n\n\n\n
mcpServers: {<\/p>\n\n\n\n
localMath: {<\/p>\n\n\n\n
command: “python”,<\/p>\n\n\n\n
args: [“.\/math_server.py”],<\/p>\n\n\n\n
transport: “stdio”,<\/p>\n\n\n\n
},<\/p>\n\n\n\n
composio: {<\/p>\n\n\n\n
url: “https:\/\/mcp.composio.dev\/your-instance\/sse”,<\/p>\n\n\n\n
transport: “streamable_http”,<\/p>\n\n\n\n
auth: { apiKey: COMPOSIO_API_KEY },<\/p>\n\n\n\n
},<\/p>\n\n\n\n
myHostedTools: {<\/p>\n\n\n\n
url: “https:\/\/myhost.com\/mcp”,<\/p>\n\n\n\n
transport: “streamable_http”,<\/p>\n\n\n\n
}<\/p>\n\n\n\n
}<\/p>\n\n\n\n
LangGraph\u2019s agent will discover tools from all<\/strong> these servers and pick the right one based on user input.<\/p>\n\n\n\nBy adding hosted and managed MCP servers, your agent now gets:<\/p>\n\n\n\n
\n- Instant access to real-world tools<\/strong> like Gmail, GitHub, Slack, Notion, etc.<\/li>\n\n\n\n
- Secure authentication<\/strong> through API keys, OAuth, or custom headers<\/li>\n\n\n\n
- Seamless integration<\/strong> of your custom or local tools alongside third-party ones<\/li>\n<\/ul>\n\n\n\n
Now that you\u2019ve got a mix of tools, it\u2019s time to combine them. Let\u2019s make sure your agent can use both local and remote tools together.<\/p>\n\n\n\n
How do you combine multiple transports or servers?<\/h2>\n\n\n\n
You\u2019ve built your tools. You\u2019ve connected your client. Now comes the fun part = mixing transports<\/strong> so your agent can talk to both local and hosted tools<\/em> at the same time.<\/p>\n\n\n\nNo matter if you\u2019ve got a math tool running locally or a Slack integration hosted on the cloud, LangGraph\u2019s MCP client can handle it all in one unified setup.<\/p>\n\n\n\n
Let\u2019s see how to do it in Python and JavaScript.<\/p>\n\n\n\n
Why combine transports? Simple: because your tools live in different places.<\/p>\n\n\n\n
\n- You might want local tools<\/strong> (like math operations or file access) for quick responses and testing.<\/li>\n\n\n\n
- And you might also need hosted services<\/strong> (like Gmail, Trello, or weather APIs) that live on the internet and require authentication.<\/li>\n<\/ul>\n\n\n\n
By combining transports (stdio, http, and sse) you let your agent choose from all<\/em> available tools, no matter where they\u2019re running.<\/p>\n\n\n\nLet\u2019s make it visual. Here\u2019s what a real-world MCP setup looks like:<\/p>\n\n\n\n![\"unified](\"https:\/\/lh7-rt.googleusercontent.com\/docsz\/AD_4nXdId001UG63ZS3iUzIDuaYzXMf7zQpi_8rBgnd-fBi7sz8Hju1cr5ybAktj_7zkEVclep53TFyX7btJQlaehWRMlaLEYYdOwqqo9Y3fVIjH486vcU5LZPkcRE7c735qTKImTIS0?key=HWm2Uf7Y276CZtzB17o9eg\")
<\/figure>\n\n\n\nLet\u2019s wire it up.<\/p>\n\n\n\n
Python: Mix Stdio and HTTP in one client<\/h3>\n\n\n\n
In Python, it\u2019s as simple as passing a config dictionary to MultiServerMCPClient.<\/p>\n\n\n\n
from langchain_mcp_adapters.client import MultiServerMCPClient<\/p>\n\n\n\n
client = MultiServerMCPClient({<\/p>\n\n\n\n
“math”: {<\/p>\n\n\n\n
“command”: “python”,<\/p>\n\n\n\n
“args”: [“.\/math_server.py”],<\/p>\n\n\n\n
“transport”: “stdio”,<\/p>\n\n\n\n
},<\/p>\n\n\n\n
“composio-tools”: {<\/p>\n\n\n\n
“url”: “https:\/\/mcp.composio.dev\/…\/server?transport=sse”,<\/p>\n\n\n\n
“transport”: “streamable_http”,<\/p>\n\n\n\n
“headers”: {<\/p>\n\n\n\n
“Authorization”: “Bearer YOUR_KEY”<\/p>\n\n\n\n
},<\/p>\n\n\n\n
}<\/p>\n\n\n\n
})<\/p>\n\n\n\n
tools = await client.get_tools()<\/p>\n\n\n\n
In this example:<\/p>\n\n\n\n
\n- The math tool runs locally through stdio<\/strong>.<\/li>\n\n\n\n
- The composio-tools endpoint connects to hosted services like Slack, Gmail, or GitHub<\/strong> via streamable HTTP (SSE)<\/strong>.<\/li>\n<\/ul>\n\n\n\n
You don\u2019t need to manage any connections manually = the client handles it for you. Tools from both sources show up in one neat list, ready for your agent to use.<\/p>\n\n\n\n
JavaScript\/TypeScript: Same idea, different syntax<\/h3>\n\n\n\n
If you\u2019re working in Node or a front-end app, the JS client offers the same flexibility, with just slightly different setup.<\/p>\n\n\n\n
Here\u2019s how to mix local and hosted transports:<\/p>\n\n\n\n
import { MultiServerMCPClient } from “@langchain\/mcp-adapters”;<\/p>\n\n\n\n
const client = new MultiServerMCPClient();<\/p>\n\n\n\n
\/\/ Connect to a local tool (stdio)<\/p>\n\n\n\n
await client.connectToServerViaStdio(<\/p>\n\n\n\n
“math”,<\/p>\n\n\n\n
“python”,<\/p>\n\n\n\n
[“.\/math_server.py”]<\/p>\n\n\n\n
);<\/p>\n\n\n\n
\/\/ Connect to a hosted tool (Composio via SSE)<\/p>\n\n\n\n
await client.connectToServerViaSSE(<\/p>\n\n\n\n
“composio”,<\/p>\n\n\n\n
“https:\/\/mcp.composio.dev\/…\/server?transport=sse”,<\/p>\n\n\n\n
{ Authorization: “Bearer YOUR_KEY” },<\/p>\n\n\n\n
true<\/p>\n\n\n\n
);<\/p>\n\n\n\n
\/\/ Load tools<\/p>\n\n\n\n
const tools = await client.getTools();<\/p>\n\n\n\n
That\u2019s it. You\u2019re now combining:<\/p>\n\n\n\n
\n- A local math server<\/strong> running on your machine<\/li>\n\n\n\n
- A hosted Composio endpoint<\/strong> with 250+ ready-to-use services<\/li>\n<\/ul>\n\n\n\n
All tools are merged under the hood and made available to your LangGraph or LangChain agent automatically.<\/p>\n\n\n\n
With everything connected, let\u2019s explore how this works in practice. You\u2019ll see how real agents use MCP tools in everyday tasks.<\/p>\n\n\n\n
What are some real\u2011world use cases?<\/h2>\n\n\n\n
Now that your LangGraph agent is connected to MCP tools, let\u2019s see what it can actually do<\/em>.<\/p>\n\n\n\nThis cheat sheet gives you inspiration to build your next MCP tool:<\/p>\n\n\n\nUse case<\/strong><\/td>LangGraph MCP example<\/strong><\/td>Tool type<\/strong><\/td>Transport<\/strong><\/td><\/tr>| Do quick math<\/td> | add(a: int, b: int)<\/td> | Local Python tool<\/td> | stdio<\/td><\/tr> | | Convert currency<\/td> | convert_currency(from: str, to: str, amount: float)<\/td> | External API wrapper<\/td> | http or sse<\/td><\/tr> | | Check the weather<\/td> | get_weather(city: str)<\/td> | Hosted API tool<\/td> | streamable_http<\/td><\/tr> | | Get current time in a timezone<\/td> | get_time(timezone: str)<\/td> | Local utility<\/td> | stdio<\/td><\/tr> | | Generate a strong password<\/td> | generate_password(length: int)<\/td> | Utility\/microservice<\/td> | stdio<\/td><\/tr> | | Summarize an email thread<\/td> | summarize_email(thread_id: str)<\/td> | Gmail (via Composio)<\/td> | streamable_http<\/td><\/tr> | | Send a calendar invite<\/td> | create_calendar_event(title, time, invitees)<\/td> | Google Calendar (Composio)<\/td> | streamable_http<\/td><\/tr> | | Create a GitHub issue<\/td> | create_github_issue(repo, title, body)<\/td> | GitHub (Composio)<\/td> | streamable_http<\/td><\/tr> | | Get a GitHub repo\u2019s stars<\/td> | get_stars(repo: str)<\/td> | GitHub API<\/td> | http<\/td><\/tr> | | Post a Slack message<\/td> | send_slack_message(channel, text)<\/td> | Slack (Composio)<\/td> | streamable_http<\/td><\/tr> | | Create a Trello card<\/td> | create_task(title: str)<\/td> | Trello (Composio)<\/td> | streamable_http<\/td><\/tr> | | Search a local folder<\/td> | search_files(query: str)<\/td> | Local Python tool<\/td> | stdio<\/td><\/tr> | | Read text from PDF<\/td> | extract_pdf_text(file_path: str)<\/td> | Local file utility<\/td> | stdio<\/td><\/tr> | | Store and retrieve user notes<\/td> | save_note(title, content) \/ get_notes()<\/td> | Custom note service<\/td> | http or stdio<\/td><\/tr> | | Translate text<\/td> | translate(text: str, to_lang: str)<\/td> | Hosted API (e.g. DeepL)<\/td> | streamable_http<\/td><\/tr> | | Trim and clean up text<\/td> | clean_text(input: str)<\/td> | String utility<\/td> | stdio<\/td><\/tr> | | Count words or characters<\/td> | count_words(text: str)<\/td> | Local tool<\/td> | stdio<\/td><\/tr> | | Convert to uppercase\/lowercase<\/td> | to_uppercase(s: str) \/ to_lowercase(s: str)<\/td> | String utility<\/td> | stdio<\/td><\/tr> | | Generate blog outline<\/td> | generate_outline(topic: str)<\/td> | LLM-powered tool<\/td> | http (with OpenAI)<\/td><\/tr> | | Answer FAQ from docs<\/td> | search_docs(query: str)<\/td> | RAG \/ document retriever<\/td> | http or stdio<\/td><\/tr> | | Add product to cart<\/td> | add_item(name: str, quantity: int)<\/td> | E-commerce logic<\/td> | stdio<\/td><\/tr> | | Show current cart<\/td> | get_cart()<\/td> | E-commerce logic<\/td> | stdio or http<\/td><\/tr> | | Remove item from cart<\/td> | remove_item(name: str)<\/td> | E-commerce logic<\/td> | stdio<\/td><\/tr> | | Get stock market data<\/td> | get_stock_price(ticker: str)<\/td> | Finance API wrapper<\/td> | http or streamable_http<\/td><\/tr> | | Send email reply<\/td> | reply_email(thread_id, message)<\/td> | Gmail (Composio)<\/td> | streamable_http<\/td><\/tr> | | List upcoming meetings<\/td> | get_calendar_events(date_range)<\/td> | Google\/Outlook Calendar<\/td> | streamable_http<\/td><\/tr> | | Create support ticket<\/td> | create_support_ticket(subject, details)<\/td> | Helpdesk integration<\/td> | http<\/td><\/tr> | | Trigger build in CI\/CD pipeline<\/td> | trigger_build(project_id: str)<\/td> | DevOps integration<\/td> | http<\/td><\/tr> | | Fetch analytics metrics<\/td> | get_analytics(metric: str, date_range: str)<\/td> | Custom backend or SaaS<\/td> | http or sse<\/td><\/tr> | | Generate daily summary<\/td> | daily_summary(user_id: str)<\/td> | Aggregator microservice<\/td> | http or stdio<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n All of this works through LangGraph MCP, which means your agent doesn\u2019t need direct access to APIs or auth logic = Composio handles it. You just connect the endpoint.<\/p>\n\n\n\n Before going too far, it\u2019s smart to test your setup. Let\u2019s check that everything works as expected, and fix anything that doesn\u2019t.<\/p>\n\n\n\n How do you test and debug it?<\/h2>\n\n\n\nYou\u2019ve connected your MCP servers and built a client. Nice work! Now it\u2019s time to make sure everything actually works.<\/p>\n\n\n\n Testing and debugging your LangGraph + MCP setup doesn\u2019t have to be painful. Just follow a few simple steps to catch problems early and keep your tools running smoothly.<\/p>\n\n\n\n Start by checking get_tools()<\/h3>\n\n\n\nBefore you run any agents or workflows, make sure your client can see<\/em> your tools.<\/p>\n\n\n\nIn Python:<\/p>\n\n\n\n tools = await client.get_tools()<\/p>\n\n\n\n print(tools)<\/p>\n\n\n\n You should see a list of available tools, each with:<\/p>\n\n\n\n \n- The correct name<\/li>\n\n\n\n
- Inputs and types (e.g. a: int, b: int)<\/li>\n\n\n\n
- Return type (e.g. int, str, etc.)<\/li>\n<\/ul>\n\n\n\n
If a tool doesn\u2019t appear or its schema looks wrong, double-check your server code. Start simple. This call confirms the basic connection is working.<\/p>\n\n\n\n Test tool calls one at a time<\/h3>\n\n\n\nDon\u2019t jump into full agent flows just yet. First, call your tools manually to make sure they behave as expected.<\/p>\n\n\n\n Example: <\/p>\n\n\n\n result = await client.call_tool(“add”, {“a”: 3, “b”: 4})<\/p>\n\n\n\n print(result) # Should return 7<\/p>\n\n\n\n This helps you catch:<\/p>\n\n\n\n \n- Type errors (e.g. passing a string instead of a number)<\/li>\n\n\n\n
- Schema mismatches<\/li>\n\n\n\n
- Unexpected tool logic<\/li>\n<\/ul>\n\n\n\n
Get these right now, and you\u2019ll save time later when agents start using the tools automatically.<\/p>\n\n\n\n Check logs on both sides<\/h3>\n\n\n\nDebugging gets easier when you can see<\/em> what\u2019s going on.<\/p>\n\n\n\nHere\u2019s where to look:<\/p>\n\n\n\n On the server side:<\/strong><\/p>\n\n\n\n\n- Watch for startup messages<\/li>\n\n\n\n
- See incoming requests<\/li>\n\n\n\n
- Track tool errors and exceptions<\/li>\n<\/ul>\n\n\n\n
On the client side:<\/strong><\/p>\n\n\n\n | | | | |
![\"langgraph](\"https:\/\/lh7-rt.googleusercontent.com\/docsz\/AD_4nXfaldkpWo29hKVoF_M7dZjrvjmODMPLDnt7jPMIj-q87SAmDa7Ant965JZlm61OWTno1f0UhB-xd79n3pAHniyusAGB1cV18jJlt4qrTWJGhAQgJMw0vhYUzX4c_Vwbn3KbefEERQ?key=HWm2Uf7Y276CZtzB17o9eg\")
<\/figure>\n\n\n\n![\"langgraph](\"https:\/\/lh7-rt.googleusercontent.com\/docsz\/AD_4nXfU5Bm5DIaV-MaXl85U_4hyUzfOis7qkx_7jmmT7-8zL9frw1FZUoNmAHPC9UYGW59Q16zHERiUB6GzFQKPxb0X-VwlZdrEPHiQAd67diKMHP06nbJqvNno5hWCxo4M05RY3m4c?key=HWm2Uf7Y276CZtzB17o9eg\")
<\/figure>\n\n\n\n