Back to blog

Strands Agents: give your agent YouTube, TikTok and Instagram transcripts over MCP

Connect a Strands agent to TranscriptFetch's hosted MCP server with MCPClient: no key for a test, an API key for real use. Tested on strands-agents 1.65.0 with mcp 2.3.0 and 1.30.0.

TL;DR: Strands Agents can use any remote MCP server through MCPClient. Point it at https://transcriptfetch.com/mcp/try to test with no account, or at https://transcriptfetch.com/mcp with an API key in the Authorization header. Your agent then gets a get_transcript tool that works for YouTube, TikTok and Instagram links, and the fetch runs on our network, not your server's.

TranscriptFetch is our product, so weigh this guide with that in mind. We ran every snippet below on 18 November 2026 with strands-agents 1.65.0, once with mcp 2.3.0 and once with mcp 1.30.0. The two mcp versions name the transport function differently, so both are shown.

Getting started

Install

bash
pip install strands-agents mcp

Check which mcp you got, because the import changes at 2.0:

bash
pip show mcp | grep Version
Guide

Try it with no key

The trial endpoint needs no account. It gives 10 calls a day and two tools, get_transcript and search_videos, captions only.

With mcp 2.x:

python
from mcp.client.streamable_http import streamable_http_client
from strands import Agent
from strands.tools.mcp import MCPClient

transcripts = MCPClient(
    lambda: streamable_http_client("https://transcriptfetch.com/mcp/try")
)

with transcripts:
    agent = Agent(tools=transcripts.list_tools_sync())
    print(agent.tool_names)  # ['get_transcript', 'search_videos']
    agent("Summarize this video in five bullet points: https://youtu.be/dQw4w9WgXcQ")

With mcp 1.x the function is streamablehttp_client, with no underscore between the words:

python
from mcp.client.streamable_http import streamablehttp_client

transcripts = MCPClient(
    lambda: streamablehttp_client("https://transcriptfetch.com/mcp/try")
)

Strands also accepts the client itself in tools=[transcripts] and opens and closes the connection for you. In our test that form registered the same two tools.

Use your API key

For regular use, connect https://transcriptfetch.com/mcp and send your key as a Bearer token. A free account starts with 50 credits and no card. A caption transcript costs 1 credit, and failed requests are free. Keys are on the API keys page.

With mcp 2.x the headers go on the HTTP client:

python
import os
from mcp.client.streamable_http import streamable_http_client, create_mcp_http_client
from strands import Agent
from strands.tools.mcp import MCPClient

key = os.environ["TRANSCRIPTFETCH_API_KEY"]

transcripts = MCPClient(
    lambda: streamable_http_client(
        "https://transcriptfetch.com/mcp",
        http_client=create_mcp_http_client(headers={"Authorization": f"Bearer {key}"}),
    )
)

With mcp 1.x they are a keyword argument:

python
transcripts = MCPClient(
    lambda: streamablehttp_client(
        "https://transcriptfetch.com/mcp",
        headers={"Authorization": f"Bearer {key}"},
    )
)

We checked both forms against a local server, and each sent the Authorization header as written. A missing or wrong key comes back as an Unauthorized error, which Strands raises as MCPClientInitializationError when the connection opens.

With a key the agent gets twelve tools instead of two. Besides the transcript and search tools there are tools to list a channel's, a profile's or a playlist's videos, to check your credits, and to set up monitors that watch a channel for new uploads. Videos with no captions can be transcribed with AI, which the trial does not do. The MCP docs describe every tool.

Call a tool without the model

To check the connection before you spend model tokens, call the tool directly:

python
with transcripts:
    agent = Agent(tools=transcripts.list_tools_sync())
    result = agent.tool.get_transcript(video="dQw4w9WgXcQ", timestamps=False)
    print(result["status"])              # success
    print(result["content"][0]["text"][:200])

timestamps=False leaves out the timed segments and keeps the response smaller, which matters when the transcript goes into the model's context. markdown=True returns paragraphs that each start with a time linked to that moment in the video, which is useful when the agent has to cite. chunks=True adds pieces of about 500 tokens for an embedding model.

What we did not test

We did not run a full conversation with a model, because the sandbox we test in has no model credentials. Strands uses Amazon Bedrock by default, and the model you choose has to support tool calls. The tool registration and the direct tool calls above are what we ran.

FAQ

Yes. The transcript is fetched by TranscriptFetch's servers, so YouTube never sees your function's address. That is the usual reason youtube-transcript-api fails on AWS, and our production guide covers it.