Post to X (Twitter) from Claude or ChatGPT with an MCP server

Is there an official X MCP server?

X runs three. The hosted server at api.x.com/mcp, live since June 30, 2026, covers search, users, timelines, bookmarks, trends and Articles, and cannot create ordinary posts. xmcp, which you host yourself, turns every endpoint of the X API into a tool, posting included, using your own X developer app. The third, X Ads MCP, is for advertising. X API access is sold as pay-per-use credits: creating a post costs $0.015, or $0.20 when it contains a link, billed to whoever owns the app.

Read on 1 October 2026 from each project’s own page. Descriptions of other servers are their own claims; Mellow has not tested them.

Servers that log in with your account’s cookies to read or post break X’s terms, which forbid scraping. Since February 23, 2026, X also accepts an automated reply only when the author of the original post mentioned or quoted the replying account.

What X account do I need?

Any X account you can sign in to; Hub posts through its provider’s X app, so you need no developer account and no API credits of your own. When you connect, Hub asks for the OAuth version: OAuth 2.0 is the default. OAuth 1.0a exists for apps still on the older credentials.

1. Open Accounts in Hub and choose X, or ask the assistant to call connect_channel with platform set to x. That tool only returns a link and an instruction for a person. It needs the channels:connect permission.
2. Open the link and sign in to X yourself. Hub never sees your X password.
3. Tell the assistant you are done. It calls list_channels with refresh: true and uses the returned channel ID, which begins with spc_. A username is not a substitute for that ID.

Connect Claude, ChatGPT, Cursor or another MCP client

Every client uses the same server URL, https://www.mellow.world/mcp, over Streamable HTTP. A client signs in with OAuth 2.1 (dynamic client registration and PKCE) or sends an owner-created Hub API key as a Bearer token. Menus change between client versions, so the official guides linked below describe the current screens. The last column says only what Mellow itself has tested, on any network.

1. Add the server in your client and select Connect. Sign in to Mellow if asked. Hub’s consent screen names the app that is asking.
2. Choose the permissions the task needs (below), select your X channel, choose review or autopilot, a daily ceiling and an expiry, then approve.
3. Back in the client, run the read-only check from the prompts below. Confirm the profile, the mode and the X channel ID. Agents shows the grant’s permissions, channels and expiry.
4. If a client cannot complete OAuth, create a key in Agents and set it as the client’s Authorization: Bearer header, from its secret settings or an environment variable. Never paste a key into a chat.

Which permissions to approve

The consent screen arrives with every permission the assistant asked for already ticked, except spending AI credits. Untick anything it should not have.

What you delegate, and how to take it back

Hub treats an assistant’s access as a delegation you give once, with limits: the permissions, the channels (the ones you select, or explicitly all), review or autopilot, a daily ceiling of 1 to 1,000 posts (50 by default) and an expiry of 1, 7 or 30 days (30 by default). Autopilot is the default: the assistant publishes on its own within those limits. Choose review and each post waits for you to approve it in Hub. A key or a connected app cannot approve its own post.

Hub records the posts an assistant creates or has refused, approvals, reschedules, cancellations and connection attempts in a journal you read in Agents. Revoking a key or a connected app takes effect on its next request. Publishing always needs an idempotency key, so a retried request cannot publish twice.

X rules Hub checks before anything is sent

This table is generated from the registry that validate_post enforces, the same one behind the rules for every network. These are Hub’s input rules; X still judges the media itself when it fetches it.

280 is the limit for a standard account, counted the way X counts: a link is 23 however long it is, and an emoji or a Chinese, Japanese or Korean character is 2. A premium account may write far more, but Mellow cannot see the account's tier, so anything longer is reported rather than silently sent.

Hub counts an X post the way X does, by weight: every link is 23 however long it is, and an emoji or a Chinese, Japanese or Korean character is 2. A post Hub accepts is one a standard account can publish; over 280, give X its own caption in perChannel. The free character counter shows the same count as you type.

A poll goes in options.x.poll with two to four options and a duration of 5 to 10,080 minutes.

When validate_post refuses the post

Read ok, issues and notes in the result. The check publishes nothing and reports every problem at once, each with its field and, for a network rule, its channel. Fix them together and validate again.

Other refusals concern the request or the grant: schedule_in_past (use a future time with a UTC offset), insufficient_scope (approve a delegation with the missing permission), channel_out_of_scope (the channel is not in the grant), and at creation daily_limit_reached and idempotency_key_reused. A passing check is not proof of publication: X can still refuse the media or the account. The free post checker runs the same input rules on a draft before you connect anything.

Prompts you can paste

Replace the bracketed parts. Each prompt names the tools to call, so the assistant does not have to guess the order.

Use only Mellow Hub. Call whoami, then list_channels, then list_platforms.
Tell me my X channel ID, whether this connection is in review or autopilot
mode, how many posts it may still create today, and X's rules. Do not
create anything.

Use Mellow Hub. Announce [what] on my LinkedIn and X channels.
LinkedIn: [the full version]
X: a version of 280 characters or fewer, same point, no hashtags.

Put each caption in perChannel under its channel ID. Validate, fix every
issue, then create_post once with idempotencyKey "launch-[short-name]-1"
and get_post. Report each target's status and link.

Use Mellow Hub. Schedule a poll on my X channel for [time] on [date] in
[your timezone]: question "[question]", options [A], [B], [C],
running for 1440 minutes (options.x.poll). Write scheduledAt with an
explicit UTC offset. Validate, then create_post once with idempotencyKey
"poll-[date]", then get_post.

Media has to be reachable at a public https URL when X fetches it. An agent that can make HTTP requests, such as Claude Code or a script of your own, can instead call request_upload_url, send a local file with PUT to the returned uploadUrl and use the returned mediaUrl. An assistant that only has Hub’s tools cannot send file bytes itself.

Publish once, then read the X result

create_post is the step that publishes or schedules. It needs an idempotencyKey of 8 to 120 letters, digits or the characters . _ : -, stable for this post. If the response is lost or times out, send the same request with the same key: Hub returns the original post instead of making a second one. Reusing a key for different content is refused.

Under a review-mode grant the post comes back as pending_review and waits until you approve it, signed in, in Hub. Under autopilot it goes out at scheduledAt, or immediately without it. Always write scheduledAt with Z or an explicit UTC offset.

Then call get_post and read the X entry in targets, not only the post’s status. publishing means X has not reported a result yet; published comes with the post’s link; failed comes with the network’s message. In a post to several networks each target settles on its own and the post can end partial: republish only the failed channel, with a new key. Once X has published a post, Hub cannot edit or delete it; cancel_post works only before it goes out.

Each destination counts once against your plan: one post to X and two other networks is three publications. Validating, drafting and connecting work without a plan; publishing needs an active one. See plans.

What Mellow has tested with X

Production records on 1 October 2026 show five X publications through Hub between September 5 and October 1, each reaching published with a public link, all on one account. Two came through Hub’s MCP endpoint from a Claude Code session using an owner-created key: a video on September 29, posted together with Instagram, Facebook, Threads, LinkedIn and TikTok, and on September 30 a post that had been scheduled the evening before and went out at its set time, after Claude’s claude.ai connector had moved that time. On October 1 a post created with an owner-created key on September 30 at 15:59 UTC for 12:30 UTC went out at 12:31 UTC, together with Facebook, Instagram, LinkedIn, Pinterest and Threads, with nothing changed in between.

Not tested yet: polls, posts with four images, and connecting through ChatGPT or Cursor.

Questions

Related

• Character counter with X’s weighted count.
• LinkedIn MCP guide and Threads MCP guide.
• Publish to social media with an MCP server: the workflow for every network.
• Before an AI agent posts: an MCP checklist.
• API and MCP reference and rules for every network.
• Plans: each destination counts once against the monthly publications.