Twitter MCP server (X) in 2026: official, open source or scraper?

By Jithin Palepu, founder of First100. Published 2026-09-27. Last updated 2026-09-27.

A Twitter MCP server comes in three kinds in 2026. X's own hosted server reads posts, users, trends and bookmarks, but it has no tool that creates a post. Open-source servers run on your own X developer app and can post, at X's pay-per-use prices. Scraper-based servers read without the API, which X's Terms of Service prohibit.

Key takeaways

• X launched a hosted MCP server at api.x.com/mcp in June 2026, and TechCrunch reported on June 30 that it cannot be used to post on X.
• X's pricing page charges $0.005 per post read, $0.010 per user read, $0.015 per post created and $0.200 per post with a URL.
• The most-starred open-source server, EnesCinr/twitter-mcp, had 407 GitHub stars and two tools when we checked on 2026-09-27.
• X's developer guidelines say non-API automation such as scraping or browser automation "results in permanent suspension."
• First100 publishes and schedules X posts through the official API over OAuth, and it never replies, likes or follows on your behalf.

What is a Twitter MCP server?

A Twitter MCP server is a program that gives an AI agent tools for X through the Model Context Protocol. Claude Code, Cursor or Codex lists the tools and calls them while it works: search posts, read a profile, publish a post. What those tools can do depends on who runs the server and how it reaches X. There are three answers, and they differ on posting, cost and risk.

The first kind is X's own server, hosted by X. The second is an open-source server you run on your machine with keys from your own X developer app. The third reads X through a scraping service instead of the API.

I build First100, a hosted go-to-market server that publishes on X among other things. So I have a stake in this. I will still start with the free, official option, because for some readers it is all they need.

What does X's official MCP server do?

X's official MCP server is a hosted endpoint at https://api.x.com/mcp, documented on X's MCP page. It gives an agent X's read API as tools: full-archive search, user lookups, timelines, mentions, trends and news. It can also manage bookmarks and publish long-form Articles. It has no tool for ordinary posts, replies, likes or direct messages.

X's page lists six tool groups:

• Posts: fetch posts, see likers, reposters and quoters, and recent counts.
• Search: full-archive post search, user search and news search.
• Users: the current user, lookups by id or handle, timelines and mentions.
• Bookmarks: list, add and remove bookmarks and folders.
• News and trends: news stories and trends by location.
• Articles: create draft Articles and publish them.

There are two ways to connect. An app-only bearer token connects directly and is read-only. For user actions, X recommends a local bridge, npx @xdevplatform/xurl mcp, which runs an OAuth 2.0 login with PKCE in your browser. X's page says to allow a startup timeout of at least 300 seconds on the first run, so the login can finish. Both routes need a CLIENT_ID and CLIENT_SECRET from your own app in the X Developer Portal.

The MCP page names no separate price for the server. The calls go through your developer app, and X's API bills that app per resource. TechCrunch's Sarah Perez put the posting limit plainly: the server "is not compatible with X's Write API endpoints, so it's not possible to use it to post autonomously (or at all) on X."

X also runs a separate Ads MCP with 74 tools for campaigns, line items and audiences. That one is for advertisers, not for organic posting from your own profile.

How do open-source Twitter MCP servers work?

An open-source Twitter MCP server runs on your computer and calls the X API with keys you create. You register an app in the X Developer Portal, give it read and write permission, and paste the keys into your MCP config. The server then posts and reads as you. You pay X directly for every call, and you maintain the setup yourself.

Three examples, checked on GitHub on 2026-09-27:

• EnesCinr/twitter-mcp, MIT license, 407 stars. It has two tools, post_tweet and search_tweets, and uses an API key plus an access token and secret. Its last release, v0.2.0, dates from June 2025.
• mbelinky/x-mcp-server, MIT license, 23 stars. It adds OAuth 2.0 and image uploads, with post_tweet, search_tweets and delete_tweet.
• achetronic/twitter-mcp, Apache 2.0, 3 stars, release v0.2.4 in March 2026. It has the widest list: timelines, mentions, threads, likes, reposts, follows and a local scheduling queue kept in a schedule.yaml file.

The appeal is control. The code is readable, the keys stay on your machine, and nothing sits between you and X. The cost is the work. You create the developer app, fund its credits, rotate tokens and keep the process running. A local schedule queue only publishes while that machine and server are up.

Some of these tools also cross X's rules if an agent uses them in a loop. A like_tweet or follow_user tool is fine when you ask for one like. The guidelines forbid auto-liking "by keyword, hashtag, user, or schedule."

What are scraper-based Twitter MCP servers?

A scraper-based server reads X without the official API. It calls a scraping service that collects posts from X's pages and returns them as data. Apify's Tweet Scraper V2, for example, costs $0.40 per 1,000 tweets and plugs into agents through Apify's MCP server at mcp.apify.com. Its page showed 102,252 total users and a rating of 3.86 out of 5 in September 2026. At $0.0004 a tweet, it is 92% cheaper per post than the API's $0.005.

The price reflects the rules it sits outside of. X's Terms of Service say that "crawling or scraping the Services in any form, for any purpose without our prior written consent is expressly prohibited." The same terms set liquidated damages of $15,000 per 1,000,000 posts for anyone who requests or views more than one million posts in 24 hours in breach of them.

X's developer guidelines add the account side. They list "Browser scripting, scraping, any automation outside official API" among activities that get an app "suspended or permanently banned." Their words on the outcome: "Non-API automation (scraping, browser automation) results in permanent suspension. Always use the official X API."

Whether that risk lands on you, the scraping vendor or nobody depends on how the scraper runs. I am not a lawyer, and this is not legal advice. Keep your main X account away from anything that reads X outside the API.

What does the X API cost an MCP server?

Any server that uses the X API, official or open source, runs on X's pay-per-use pricing. You buy credits and X deducts them per resource returned or per request. Reads are cheap one at a time and add up fast in bulk. A post with a link costs about 13 times a plain post in 2026, at $0.200 against $0.015.

The prices below are from X's pricing page, checked on 2026-09-27.

Two rules soften this. The same resource is free again within one 24-hour UTC day, because X deduplicates it. Pay-per-use plans stop at 3 million post reads per billing month; beyond that X wants an Enterprise plan. X also gives up to 20% back in xAI API credits once spend reaches $1,000 in a cycle.

A search that returns 100 posts costs $0.50. Ten link posts a week cost $2.00. Plan the agent's reads before you give it a search tool and a loose instruction.

How do the three kinds compare?

The three kinds trade setup work, posting ability and account risk against each other. X's server is the easiest way to read X with your account, but it does not post. Open-source servers post and read on your own keys. Scrapers read cheaply, outside X's terms. The table sets them side by side, with First100 as a fourth column for comparison.

Where does First100 fit?

First100 is a hosted MCP server for go-to-market work, and X is one of its four platforms. It publishes and schedules X posts through X's official API over OAuth, so you need no developer app of your own. Your agent writes each post. First100 stores it, checks it, publishes it at its time and measures it afterwards.

What it adds around a plain post tool:

• A scheduler with retries. A job runs every minute, and a failed publish retries after 5, 10 and 20 minutes. Identical text on X within 24 hours is rejected, and a platform gets at most 25 posts a day.
• X's own counting. 280 characters, every link counted as 23, emoji and CJK counted as 2. Images up to 5 MB in JPEG, PNG, GIF or WebP.
• Brand memory. get_content_context hands the agent your audience, tone, pillars, product updates and recent posts before it writes.
• A pre-flight check. review_post flags a weak hook, extra links, filler and engagement bait before a draft is saved.
• Tracked links. URLs in a post become /l/ links, so you see clicks per post and per platform.
• Reply radar. reply_radar watches X accounts you choose and ranks their new posts as reply chances. It drafts; you post the reply yourself.

The X integration page lists these limits in one place.

What can't First100 do on X?

First100 has the permissions an ordinary X app gets, and it chooses not to use some of them. It publishes, schedules and measures your own posts. It does not message anyone, engage for you, or give you full-archive search the way X's official server does. Here is the list, so nobody finds out after paying.

• No direct messages. It cannot read or send DMs.
• No automated replies, likes or follows, by design. X's guidelines forbid unsolicited automated replies and auto-liking, and require approval for AI-generated replies. The radar only surfaces and drafts.
• Metered reads. Metrics and radar scans count against a monthly X budget: $4 on the week trial, $10 on Starter and $10 on Pro. The radar watches 3, 8 or 15 X accounts. Publishing never counts against that budget.
• One X account per brand. Starter covers 2 brands and Pro covers 5, so a second X account needs a second brand.
• No general search tool. For searching all of X, use X's own server next to First100.

How do you post to X from Claude Code or Cursor with First100?

Connecting takes a few minutes and needs no X developer app. You connect X in the dashboard, add First100's MCP endpoint to your agent, and ask for a post. The steps below cover Claude Code and Cursor. The quickstart guide covers Codex, ChatGPT and Claude Desktop.

1. Create your account. Sign in at firsthundred.app and pick a plan. The week trial costs $5 for 7 days with 20 posts; Starter is $39 a month with 300 posts, and Pro is $59 a month with 600.

2. Connect X. Open Connected accounts in the dashboard, choose X, sign in and approve. First100 asks for tweet.read, tweet.write, users.read, media.write and offline.access.

3. Add First100 to your agent. In Claude Code, run:

   claude mcp add --transport http first100 https://firsthundred.app/api/mcp \
     --header "Authorization: Bearer f100_your_key"
   

   In Cursor, put the same URL and header in ~/.cursor/mcp.json.

4. Check the connection. Ask: "Use First100 to list my connected accounts." X should appear with your handle.

5. Ask for a post. For example: "Read my brand context, write an X post about this week's release, review it, and schedule it for Tuesday at 9:00 in my timezone." The agent calls get_content_context, writes, runs review_post, then save_drafts.

6. Watch the queue. The post shows in the dashboard calendar. With auto-publish on, the default, it goes out at its time.

7. Read the numbers later. After a few days, ask for get_post_performance or what_worked to see impressions, replies and reposts per post.

Which Twitter MCP server should you use?

Pick by the job. If you only want your agent to read X, search it and look up users, start with X's official server; it costs nothing beyond X's API prices. If you want to post from a script you control, an open-source server on your own developer app does it. If you want scheduled posts with checks, measurement and no app to maintain, that is the gap First100 fills.

Three questions settle it:

1. Do you need to post, or only to read? Posting rules out X's own server today.
2. Do you want to run infrastructure? Open-source servers need a developer app, credits and a running process.
3. Is your X account an asset you cannot lose? Then stay on the official API and keep engagement manual.

The two can also run together. X's server for research, First100 for publishing, and you for replies.

For LinkedIn, the same choice looks different; see our LinkedIn MCP server guide.

FAQ

Is there an official Twitter MCP server?

Yes. X hosts one at api.x.com/mcp, launched in June 2026. It covers search, users, timelines, trends, bookmarks and Articles. It cannot create ordinary posts, replies or likes, and it needs your own X developer app credentials.

Can a Twitter MCP server post tweets?

Open-source servers such as EnesCinr/twitter-mcp can, through your developer app's keys. X's official server cannot. First100 can, with scheduling, through the official API over OAuth.

How much does it cost to post to X through an MCP server?

X charges $0.015 per post and $0.200 per post containing a URL, according to its pricing page. On a self-hosted server you pay that from your app's credits. On First100 publishing is included in the plan.

Can a Twitter MCP server get my account suspended?

A scraper or browser-automation server can. X's developer guidelines say non-API automation "results in permanent suspension." Official-API servers carry little account risk, as long as the agent does not auto-like, auto-follow or send unsolicited replies.

Does First100 reply to tweets automatically?

No. reply_radar ranks posts worth replying to and can draft a reply, but you post it yourself. X's guidelines require approval for AI-generated replies, and First100 does not automate replies, likes, follows or DMs.

Connect X and schedule your first post

Open the quickstart guide, connect X under Connected accounts, and add the endpoint to Claude Code or Cursor. Then ask your agent for one X post about something you shipped this week, and schedule it for your next weekday morning.