The question has a yes in it, so most guides stop there. The interesting part is which server can actually do it, and what arrives on the bill afterwards.

Can Claude post to Twitter?

Yes, through a connector, but not through the server most people install. X's general-purpose hosted MCP server can read posts, search, manage bookmarks and publish Articles. It cannot create an ordinary post. Posting requires either X's self-hosted server, its newer advertising server, or a third-party one.

X's three official MCP servers, checked at its own documentation on 9 September 2026

That split is the single most misreported thing in this topic. Press coverage and several guides describe the hosted server as exposing more than two hundred endpoints including post creation. That number belongs to a different server, and the tool that creates posts is not in the hosted one.

The proof, which takes one request

Capability tables are marketing surface and can lag. There is a harder artefact, and it settles the question.

Every MCP server that uses OAuth publishes a metadata document listing the scopes it can be granted. X's hosted server publishes one, and the list includes reading posts, reading users, reading and writing bookmarks, reading and writing lists and blocks, and billing. It does not include write access to posts. The neighbouring document for the wider API does advertise media write and advertising write, so the omission is specific rather than an oversight in how the file was generated.

‼️ That is a stronger check than reading a feature list, and you can repeat it on any MCP server before you connect it. Fetch the server's protected-resource metadata and read the scopes. A server cannot do a thing it cannot be granted permission for, whatever its description says.

X also told a reporter the same thing at launch, saying the tool is not compatible with its write endpoints so it is not possible to post with it. The capability table, the scope list and the company's own statement all agree.

The server nobody noticed

Around 21 August 2026, X launched a third MCP server aimed at advertisers. It has 23 tools, authenticates natively without the bridge the general server needs, and ships documented setup instructions for Claude.

Two of those tools write posts. More interestingly, X's own reference for the endpoint behind one of them states that both promoted-only and organic post creation are supported, and that organic posts can be created for the authenticated user. So the official server that can publish an ordinary post today is the advertising one.

The trade press covering that launch described it as connecting ad data to AI tools for guidance rather than creating anything. X's own reference documents ten state-changing tools. Where those two disagree, we are going with the documentation.

One honest limit on that finding. Whether the advertising server actually exposes the organic-versus-promoted switch to the model is not published, and confirming it needs an account with advertising access. We could not test it. What is documented is that the endpoint supports both.

Now the bill

Every guide that mentions cost quotes the same number, and it is the least useful figure available. Here is the arithmetic instead, using X's published rates.

What one published post costs, from X's own rate card

Three of those rows deserve explaining, because each contains something people get wrong.

• A link costs more than thirteen times a plain post. Same single call, same endpoint, 200 thousandths of a dollar instead of 15. Your own domain gets no exemption.

• Reading twenty posts is one request and twenty charges. X bills per resource returned, not per call. That one sentence resolves most confusion about this pricing model. The same twenty posts belonging to somebody else cost five times more, because reading your own data is discounted.

• The image row is a range because X has not published enough to make it a number. Media upload calls have no price on the rate card. Each attached media object bills as an extra post creation, according to a staff answer that is absent from the pricing page. And whether that surcharge bills at the plain rate or the link rate is undocumented.

What an assistant actually costs

The sticker price assumes one call. An assistant does not work that way. It checks which account it is on, reads recent posts so it does not repeat you, drafts, posts, and confirms the thing landed.

One published post, through an assistant, using X's published rates

That is roughly two and a half times the sticker price for a plain post. Add a link and a search step, which is what a research-then-post workflow looks like, and the same cycle comes to about $0.275. That is eighteen times the number every guide quotes, for one published post.

The reassuring half: the model's thinking, drafting and rewriting cost nothing on X's meter. Failed requests are not billed, because nothing is returned. And repeats of the same resource inside a 24 hour window are deduplicated, which X describes as a soft guarantee. So polling is cheaper than the arithmetic suggests.

At a post a day, a plain cadence is a little over a dollar a month. The same cadence with a link every time is over eight. The links, not the assistant, are the bill. We went through the whole rate card in what the X API actually costs in 2026.

What happens when the money runs out

It stops. Not just writing, everything. X's documentation says requests are blocked until you add credit, and the wording is API requests rather than write requests.

The specific failure we measured on our own account is an HTTP 402 with a credits-depleted problem type. We are flagging that as our own observation rather than a citable fact, because searching X's entire published documentation corpus for that status code returns nothing. The error that stops your product is the one X has not documented.

‼️ The trap underneath it: auto-recharge is documented as paused at zero or negative balance. The feature that exists to prevent an outage is disabled by the outage. X even warns that bursty usage can produce out-of-credit errors with auto-recharge switched on. If you drain the balance faster than the five-minute top-up window allows, you stay dark until a person intervenes.

Three claims worth disbelieving

1. That new developers get $500 in free credits. Several 2026 guides say this. It traces to a hand-picked closed beta in October 2025. X's current documentation promises no signup credit at all.

2. That the hosted server exposes 200 or more endpoints including posting. That figure describes the self-hosted repository, whose README actually lists 119 tools. The hosted server publishes six read-mostly categories.

3. That the X API has a $200 a month floor. That was the Basic tier, which closed to new developers in February 2026 and was force-migrated out during 2026. There is no subscription now, only credits.

So what should you actually use?

1. If you want reading, search and bookmarks, the hosted server is the least effort and the safest install. It is X's own, and it cannot post, which for many people is a feature.

2. If you want posting and you are comfortable running something locally, the self-hosted repository does it. Note before you adopt it that it carries no licence file, which is a legal question separate from whether it works.

3. If you are running ads, the advertising server is newer, better authenticated and can publish.

4. If you want none of that, a third-party connector or a scheduling product is doing the same API calls with the same charges, just billed through somebody else.

5. Whatever you pick, set a spending cap before your first post rather than after your first surprise.

Where we sit

Our MCP connector publishes to Twitter through the official API, so these rates are ours too. We pay the link surcharge on exactly the posts creators most want scheduled, which is a real cost of running this kind of product and not one we can engineer away.

The LinkedIn side of the same question has the opposite shape. Nothing to pay, and a clause in the developer terms saying you should not automate posting at all. We wrote that up in can Claude post to LinkedIn, and the connector lifespan problem in which LinkedIn MCP servers stop working after 60 days.

If you would rather Claude drafted and you posted by hand, our Claude skills hold no credentials and call no API, so they cost nothing to run and cannot generate a bill on anyone's meter.

The honest summary is that yes is the easy half of the answer. The useful half is that the official server most people reach for cannot do it, and the number everyone quotes is roughly a tenth of what a working setup costs.