Best Social Media MCP Servers for Claude, Cursor and Codex

November 8, 2026

An MCP server gives an AI agent real tools instead of guesses. Ask Claude Code for a creator's follower count without one and it either invents a number or tries to open a page that asks for a login. Connect a social media MCP server and the agent calls a tool, gets JSON back, and can cite it.

Only a handful of MCP servers read social platforms well. This post compares the ones worth trying in Claude Code, Cursor and Codex, says what each is good at, and lists the checks to run before you give any of them an API key.

What to check before you install one

• Remote or local. A remote server is a URL plus a header. A local one runs on your machine through npx, uvx or Docker, and you maintain it.
• Read or write. Some servers can post, like or bookmark. For research, a read-only server means an agent mistake cannot publish anything.
• Coverage. One platform or many. Five single-platform servers means five keys and five output shapes.
• Output shape. If TikTok returns one field name for likes and Instagram another, the agent spends tokens translating. Consistent field names make cross-platform prompts work on the first try.
• Billing. Per call, per record, per GB or a subscription. Check whether failed calls cost money, because agents retry.
• Auth. OAuth with your own account, an app key, or a vendor key. OAuth tools act as you, with your permissions.

The options at a glance

X's official MCP server

X hosts an MCP server at https://api.x.com/mcp. You connect through xurl, a small bridge that runs with npx and handles OAuth. You need an X developer app with OAuth 2.0, and the app has to be on the pay-per-use package. The tools cover post lookups, full-archive search, user lookups, timelines, mentions, bookmarks, news and trends, and it can create and publish Articles. Reads are billed at X's API rates, which today start at $0.005 per post read (X pricing). If you only care about X and want first-party data, this is the most direct route. The trade-off is setup: an OAuth app, a browser login on first use, and a startup timeout long enough for that login to finish.

Bright Data

Bright Data offers a fully managed remote MCP or a local deployment. Its social media page names X, Instagram and LinkedIn among 100+ domains with structured data, and lists a free tier of 5,000 requests a month, then $1.5 per 1,000 requests on pay as you go. It is a good fit when the same agent also needs Amazon listings or arbitrary websites. For a longer look, see our Bright Data comparison.

Firecrawl

Firecrawl's MCP server is MIT licensed and also has a hosted endpoint at mcp.firecrawl.dev. It is built for websites: scrape, map, crawl, search and parse PDFs. Without a key only scrape, search and parse work, with rate limits. It is not a social data tool. Firecrawl's own docs say it does not scrape LinkedIn directly, and some sites return an UNSUPPORTED_SITE error. We cover the details in what Firecrawl can and cannot scrape.

Open-source single-platform servers

mcp-youtube-transcript runs with one uvx command and exposes get_transcript, get_timed_transcript, get_available_languages and get_video_info. Its README warns that YouTube blocks most automated transcript requests from cloud and data center IPs, so expect it to work on a laptop and fail on a server.

GridfireAI/reddit-mcp wraps PRAW, the Python Reddit API wrapper, with six read-only tools such as search_posts and get_comments_by_submission. You create a Reddit developer app and pass its client id and secret. Both servers are free and fine for personal research. The cost is upkeep: when the platform changes, you wait for a fix or patch it yourself.

Connect a remote server in two minutes

Claude Code takes one command:

claude mcp add --transport http scrapeio https://scrapeio.com/mcp --header "x-api-key: $SCRAPEIO_API_KEY"

Codex reads ~/.codex/config.toml:

[mcp_servers.scrapeio]
url = "https://scrapeio.com/mcp"
bearer_token_env_var = "SCRAPEIO_API_KEY"

Cursor and the other clients are on the agents page. The pattern is the same for any remote server: a URL and an auth header.

Five prompts that test a server

1. "Get the last 20 TikTok videos from @duolingo and give me the median plays. Link the 3 biggest outliers." Tests pagination and whether play counts are real numbers.
2. "Compare followers for duolingo on TikTok, Instagram and YouTube in one table." Tests cross-platform coverage.
3. "Pull the transcript of this YouTube video and list every product it names." Tests transcripts.
4. "Search X for posts about oat milk from today and group them by complaint." Tests search freshness.
5. "Ask for a handle that does not exist." Tests error handling, and whether you pay for the miss.

Count the tool calls each prompt needs. A server that takes six calls where another takes two costs more than its price sheet suggests.

Safety checklist

• Keep keys in environment variables, never in a config file you commit.
• Use read-only servers for research agents.
• Use prepaid credits or a spend cap so a loop cannot run up a bill.
• Approve tool calls by hand for the first few sessions and read what the agent sends.
• Treat text that comes back from social platforms as untrusted input. A post can contain instructions aimed at your agent.

Doing it with ScrapeIO

ScrapeIO's hosted server at https://scrapeio.com/mcp works in Claude Code, Cursor, Codex and any client that speaks Streamable HTTP. Its tools map to endpoints such as the TikTok profile API, the YouTube transcript API and the Instagram comments API. Every platform returns the same field names, it only reads public data, and failed or cached calls cost nothing. Setup for each client is on /agents, and the agent skills install with npx skills add xhoantran/scrapeio-agent.

Sources

Checked on 8 November 2026 from each vendor's own pages. Prices and limits change; follow the links before you decide.

• https://docs.x.com/tools/mcp
• https://docs.x.com/x-api/getting-started/pricing
• https://brightdata.com/ai/mcp-server/social-media
• https://github.com/firecrawl/firecrawl-mcp-server
• https://docs.firecrawl.dev/features/scrape
• https://github.com/jkawamoto/mcp-youtube-transcript
• https://github.com/GridfireAI/reddit-mcp