LangChain agents that call live trend tools through MCP

Trends MCP is the data layer. LangChain and LangGraph are agent runtimes that should bind it as tools rather than as a one-off requests wrapper per platform. The MCP client points at https://api.trendsmcp.ai/mcp with Authorization: Bearer YOUR_API_KEY and transport streamable_http. The tools the runtime must describe are get_trends, get_growth, and get_top_trends. Scheduled evaluators should skip MCP and POST REST mode: "get_time_series" (the REST name for get_trends), get_growth, and get_top_trends to https://api.trendsmcp.ai/api. Free usage is 100 requests per month. Starter is $19 per month. LangChain is the right client when trend calls must sit beside retrieval, SQL, and policy tools in a graph the team already evaluates with golden traces.

Binding MultiServerMCPClient

A typical binding looks like this:

from langchain_mcp_adapters.client import MultiServerMCPClient

client = MultiServerMCPClient({
    "trends-mcp": {
        "url": "https://api.trendsmcp.ai/mcp",
        "transport": "streamable_http",
        "headers": {"Authorization": "Bearer YOUR_API_KEY"},
    }
})

Load tools from that client and pass them into the agent or LangGraph node. Package names and import paths move across langchain-mcp-adapters versions; the URL, Bearer header, and tool names do not. Cursor uses url plus transport: "http" in ~/.cursor/mcp.json. That file is not a Python client. Windsurf’s serverUrl is not a Python client. Claude.ai’s OAuth host https://www.trendsmcp.ai/mcp is not a Python client. LangChain is a Bearer caller on the api host.

Keys come from account signup. Read them from the environment, not from a committed notebook. The Connect docs keep this snippet next to the IDE dialects so copy-paste across languages stays honest.

When an agent graph should own trend tools

LangChain wins when the product is already a graph: retrieve docs, call SQL, call trends, grade the answer, maybe loop. n8n wins when the graph is ops, not an LLM. Cursor wins when a human is driving. A second home-grown Google Trends scraper inside a Tool class recreates schema drift the MCP server already solved.

Do not add Trends MCP to a graph that only needs a daily Sheet. REST from a job is cheaper to test. Do add it when the user utterance is unbounded (“what is moving in outdoor gear”) and the planner must pick get_top_trends versus get_growth with a keyword.

Tool descriptions that survive a planner

Write descriptions that name required fields and copy legal enums. source values include google search, youtube, tiktok, reddit, amazon, npm, steam, news volume. Reddit keywords are subreddit names without r/. npm names are exact and case-sensitive. TikTok is hashtag-style. get_top_trends requires a type string such as Google Trends or GitHub Trending Repos. The data sources table is the source of truth to paste into the description, not a blog memory.

Tell the planner that percent_growth may include 3M, 12M, and YTD in one get_growth call without extra request cost, while extra sources cost extra requests. Tell it that get_trends is one source per call. Tell it to stop after a budget (for example 8 calls). Without a budget, ReAct loops will page live feeds until Free is gone.

REST evaluators should assert against get_time_series payloads so notebooks and production tools cannot drift. The same POST body belongs in pytest fixtures and in the MCP tool wrapper.

Guardrails, tracing, and split runtimes

Cache identical (mode, source, keyword, windows) tuples in Redis or a process cache for a short TTL. Log returned dates so a reviewer can trace a summary back to JSON. When growth looks extreme, force a get_trends / get_time_series follow-up before the graph is allowed to say “breakout.” Cap limit on feeds (default 25, max 200). Treat 429 as terminal for that user turn.

Many stacks already split: LangGraph for the assistant, Celery or GitHub Actions for nightly eval. Nightly eval should use REST so MCP session machinery is not a flaky dependency. Same Bearer secret, two transports.

Limits inside an LLM loop

The free plan's 100 monthly requests are a rounding error for a chatty graph with retries. Retry policies must not retry 400 invalid_source or 401. They may retry 500 once. Starter at $19 per month is the usual floor for a public agent. Pricing lists higher tiers. History is weekly over about five years; agents that promise intraday trading signals are misusing the tool.

Do not expose the API key to the browser via a LangServe playground with the secret in the client. The graph server holds the Bearer token. User-supplied source strings should be allowlisted. Pin the MCP adapter version in requirements so a silent upgrade cannot rename header fields.

LangSmith or whatever tracer the team uses should store the raw tool JSON, not only the model’s summary. A grader that checks “did the answer mention growth” will pass a hallucinated 40 percent. A grader that checks results[0].growth against the tool message will not.

Version the tool descriptions in git next to the graph. When the data sources catalog adds a feed, the planner cannot use it until the description updates. Silent catalog drift looks like model failure.

LangChain is the programmable agent runtime for the shared Trends MCP server. It is not a replacement for n8n clocks, not Claude.ai OAuth, and not an excuse for unbounded tool loops.