Which AI Agents Can Actually Post to X in 2026 (And Which Can Only Draft)

Short answer: As of September 2026, no major AI agent can post to X on its own. Every agent that publishes is doing it through either the X API with a developer account, or a third-party posting layer you connect yourself.

The qualifier that matters: the difference between agents is not whether they can post. It is how much work it takes to give them the ability, and whether they can also schedule. Most can be taught to publish in a few minutes. Almost none can schedule without help.

The Short Version

Checked September 9, 2026.

Agent	Can post to X natively?	What it actually needs	Can it schedule?
Claude (Desktop, Code, API)	No	An MCP server. MCP support has shipped since November 2024	Yes, through the MCP server
ChatGPT	No	A custom MCP connector in Developer Mode, gated to Business, Enterprise and Edu plans	Yes, through the connector
Grok Bot	No. The X connector is read-only as of August 29, 2026	A custom MCP connector at grok.com/connectors	Yes, through the connector
Perplexity Comet	No. Drafts only	Nothing works. Copy the text out, or move to Computer	No
Perplexity Computer	No	A custom MCP connector. Pro or Max, local stdio on the Mac app	Yes, through the connector
OpenClaw	No	A ClawHub plugin or a SKILL.md file	Yes
Paperclip	No	An MCP server via @paperclipai/mcp-server	Yes
Cursor	No	An MCP server in .cursor/mcp.json, or a SKILL.md	Yes
Windsurf	No	An MCP server in the Cascade MCP settings	Yes
Cline	No	An MCP server in the Cline MCP marketplace config	Yes
Codex CLI	No	An MCP server, or a SKILL.md. Reads the standard since March 2026	Yes
Google Antigravity	No	A SKILL.md. Antigravity adopted the standard in January 2026	Yes

Three more things people ask about that are not agents:

Tool	Can it post to X?	The catch
XMCP, X's own MCP server	Yes. Live at https://api.x.com/mcp since June 30, 2026, 200+ endpoints	It cannot schedule, and it bills per tool call against your X API plan
n8n, Zapier, Make	Yes, through an HTTP request node	Their native X nodes have been unreliable since the API tier changes. See our n8n 403 fix
Gemini CLI, Hermes Agent, Goose	Yes	All three read SKILL.md, so one file covers all of them

Why "Can It Post?" Is Not a Yes or No Question

People conflate three separate capabilities, and vendors are happy to let them.

1. Draft. The agent writes tweet text. Every model on the list does this, and it is not what anyone is actually asking for.
2. Publish now. The agent sends a request that makes a post appear on your account. This needs credentials, and credentials are the whole problem.
3. Schedule. The agent stores a post with a future timestamp and something publishes it later, when the agent is not running. Almost nobody has this, because it needs a server that outlives the chat session.

X's own MCP server shows the gap clearly. It exposes 200+ endpoints and it can publish, but it cannot schedule, because the X API has no scheduling endpoint on the self-serve tiers. An agent that can only publish now is an agent you have to be awake for.

The Three Ways an Agent Gets Posting Ability

1. The X API directly

This is the path most tutorials still describe, and most of those tutorials are out of date. As of April 2026, new developers cannot sign up for the Basic or Pro tiers at all. Legacy Basic at 200 dollars per month and Pro at 5,000 dollars per month continue only for existing subscribers. New accounts get metered pay-per-use or an Enterprise contract starting around 42,000 dollars per year.

Current pay-per-use rates:

• $0.015 per post created
• $0.20 per post if it contains a link
• $0.005 per post read, capped at roughly 2 to 3 million reads per billing cycle

That link surcharge is the one that catches people. An agent posting four links a day costs about 24 dollars a month in fees alone, before you have written a line of application code.

One more restriction that breaks a common agent design: since February 2026, X has banned programmatic replies on all self-serve tiers. Free, Basic, Pro and pay-per-use accounts can only reply when the original author @mentions or quotes you. Enterprise is exempt. If your plan was "an agent that replies to my mentions," read the X API rules for AI agents before you build it.

2. An MCP server

Model Context Protocol is an open standard, introduced by Anthropic in November 2024, that lets an AI client call external tools. The agent discovers what tools exist, then calls them by name. You configure the server once and every future conversation has the capability.

Two flavours matter:

• Local (stdio), launched by the client as a subprocess, usually npx -y @opentweet/mcp-server. Works in Claude Desktop, Claude Code, Cursor, Windsurf, Cline, Codex CLI and the Perplexity Mac app.
• Remote (Streamable HTTP), a URL the client connects to, for example https://mcp.opentweet.io/mcp. Works in Grok, ChatGPT Developer Mode, and any client that supports remote connectors.

The OpenTweet MCP server exposes 43 tools covering posting, scheduling, threads, X Articles, evergreen recycling, best-time analysis and analytics.

3. A SKILL.md file

The newest option, and the widest reach for the least work. SKILL.md is a cross-vendor open standard published at agentskills.io on December 18, 2025: a single markdown file with YAML frontmatter that tells an agent how to do something.

By March 2026, more than 32 tools read the same file unmodified, including Claude Code, Codex CLI, Gemini CLI, Cursor, GitHub Copilot, Google Antigravity (which adopted it in January 2026), OpenClaw, Hermes Agent, Goose, OpenHands, Amp, Trae, Roo Code, Letta, JetBrains Junie and Kiro. One file, 32 agents.

Agent by Agent

Claude (Desktop, Code, and the API)

Claude cannot reach the internet or your X account by itself. Add an MCP server and it can post, schedule, build threads up to 25 posts, batch-schedule, and read back analytics. Setup is a JSON block in the Claude Desktop config, or one claude mcp add command in Claude Code. Full walkthrough: Can Claude post tweets?

Verdict: publishes and schedules, once connected. The easiest setup on this list.

ChatGPT and Workspace Agents

ChatGPT supports custom MCP connectors, but behind a plan gate. As of September 2026 the feature lives in Developer Mode on Business, Enterprise and Edu plans. On a personal Plus account you get drafting only.

Note for anyone following an older guide: ChatGPT Atlas was discontinued on August 9, 2026, and OpenAI folded browser-based agentic work into ChatGPT and Codex. Details: Can ChatGPT post tweets?

Verdict: publishes on the right plan tier. Draft-only otherwise.

Grok and Grok Bot

This is the most counterintuitive entry on the list. Grok is owned by X, and Grok still cannot post to X out of the box.

Grok Bot launched on August 11, 2026 with scheduled routines. The X connector followed on August 29, 2026, and xAI's own announcement lists what it does: search posts, read your timeline, check mentions, pull what is happening. Publishing is not on that list, and xAI does not claim it anywhere.

The fix is that Grok supports bring-your-own MCP. At grok.com/connectors you choose New Connector, then Custom, and paste any MCP server URL reachable over the public internet. Grok auto-discovers the tools. Step by step: post to X from Grok.

Verdict: read-only natively. Publishes and schedules once you add a custom connector.

Perplexity Comet vs Perplexity Computer

These are two different products and the distinction decides the answer.

Comet is the browser. It summarises pages and drafts posts. It cannot publish to social media. You copy the text out.

Computer launched February 25, 2026 at 200 dollars per month and orchestrates 19 models. It is the agentic product and it can act. Custom MCP connectors arrived on Pro and Max in March 2026. The Mac app runs local stdio servers through Settings, then Connectors, after installing the PerplexityXPC helper. Remote HTTP connectors were still rolling out and not generally available as of July 2026, so lead with the local path. More: can Perplexity post to X?

Verdict: Comet drafts only. Computer publishes and schedules with a local MCP server.

OpenClaw

OpenClaw stood at 387,250 GitHub stars on August 23, 2026, which makes it the 14th most-starred project ever. ClawHub, its tool registry, lists roughly 52,000 tools with 180,000 users and 12 million downloads.

It has two paths to posting: install a ClawHub plugin, or drop a SKILL.md into the skills directory. Both work. The plugin is faster; the SKILL.md is portable to the other 31 agents that read the format.

Verdict: publishes and schedules. Two setup routes, both short.

Paperclip

Paperclip launched on March 2, 2026 and hit 74,000+ GitHub stars by September 2026, with the latest release v2026.831.1 on September 2, 2026. It is MIT-licensed and free, and it wraps agents in a company structure with a CEO agent that hires role-specialised sub-agents. MCP support ships through @paperclipai/mcp-server.

Verdict: publishes and schedules through MCP. The fastest-growing agent on this list.

Coding agents: Cursor, Windsurf, Cline, Codex CLI, Antigravity

All five connect the same way, and none has a native X integration. Cursor takes a .cursor/mcp.json entry. Windsurf takes an MCP entry in Cascade settings. Cline uses its MCP config. Codex CLI and Google Antigravity both read SKILL.md, so one file covers them without any MCP configuration at all.

Verdict: all five publish and schedule. Same API key, different config file.

Agent frameworks: LangGraph, CrewAI, OpenAI Agents SDK, Mastra

Frameworks are a different shape of problem. There is no config file to edit; you register a tool in code. Every one of these supports either an MCP client or a plain HTTP tool, so the integration is a function that calls POST https://opentweet.io/api/v1/posts with a Bearer key. Covered in post to X from AI agent frameworks.

Verdict: publishes and schedules. Requires writing code, unlike everything above.

Automation platforms: n8n, Zapier, Make

All three can post to X, and all three have native X nodes that have been unreliable since the API tier changes. The route that keeps working is a generic HTTP request node pointed at a posting API. If you have hit the classic n8n error, see fix n8n Twitter 403.

Verdict: publishes and schedules through an HTTP node.

X's own MCP server

XMCP went live on June 30, 2026 at https://api.x.com/mcp over Streamable HTTP, first-party, exposing more than 200 X API endpoints. It can post. Never describe it as read-only.

Three things it does not solve. It cannot schedule. It bills per tool call against your X API plan, which means the 0.20 dollar link fee applies to every link post your agent makes. And user-context writes need the xurl mcp local bridge plus an X developer app with OAuth 2.0, which is exactly the developer account most people are trying to avoid. Full breakdown: the XMCP guide and posting to X through MCP.

Verdict: publishes. Cannot schedule. Needs the developer account.

What "Autonomous" Should Actually Mean

The framing most vendor pages skip: you probably do not want an agent that publishes without you. X's automation rules require that automated accounts do not spam, do not post duplicative content, and do not artificially amplify. An agent looping unattended fails all three eventually, and the failure mode is not a warning email. It is a suspended account.

The design that works in practice:

• Give the agent schedule and draft permissions, not publish_now
• Review the queue once a day, which takes about two minutes
• Reserve immediate publishing for cases where you are in the conversation watching it happen

Every capability in the table above supports this split. The agent does the writing and the timing. You keep the final yes.

What It Costs, Per Agent

The agent is rarely the expensive part. The X connection is.

Route	Setup cost	Per-post cost	Link post	Can schedule
X API pay-per-use	Developer account, OAuth app	$0.015	$0.20	No
X API Enterprise	Contract, from roughly $42,000/year	Included	Included	No
XMCP (X's own)	Developer app plus xurl mcp bridge	Billed to your X plan	$0.20	No
OpenTweet	An ot_ API key, no developer account	Included	Included	Yes

OpenTweet is $11.99/month on Pro, $29 on Advanced, $49 on Agency, with a 7-day free trial and no credit card. Annual works out to $9.99, $24.17 and $40.83 per month. Rate limit is 60 requests per minute. See pricing.

The break-even against pay-per-use is roughly 60 link posts a month. Past that, metered X API access costs more than a flat subscription and still cannot schedule.

Two Older Guides, Narrowed

If you landed here from one of our earlier posts, here is how they now fit. AI agents that post to Twitter and how to make an AI agent post to Twitter were both written in February 2026. They predate the April 2026 X API tier closure, Perplexity Computer, Paperclip, Grok Bot and the cross-vendor adoption of SKILL.md. Read the first for a side-by-side of five setups and the second for three methods with code. For the question of which agents can post at all, this page is the current answer.

Frequently Asked Questions

Which AI agents can actually post to X in 2026, and which can only draft?

As of September 2026, no major AI agent can post to X on its own. Claude, ChatGPT, Grok, Perplexity Computer, OpenClaw, Paperclip, Cursor, Windsurf, Cline, Codex CLI and Google Antigravity can all publish once you connect a posting tool, usually an MCP server or a SKILL.md file. Perplexity Comet is draft-only and cannot publish at all.

Can Claude post tweets?

Yes, once you connect an MCP server. Claude Desktop, Claude Code and the Claude API have supported MCP since November 2024, but Claude has no built-in X connection. With the OpenTweet MCP server installed, Claude gets 43 tools and can post, schedule, thread and pull analytics. Without it, Claude can only write text for you to paste.

Can ChatGPT post to X for me?

Only with an MCP connector, and the connector feature is plan-gated. As of September 2026, custom MCP connectors run in ChatGPT Developer Mode on Business, Enterprise and Edu plans. On a personal Plus account ChatGPT drafts the post and you paste it. ChatGPT Atlas, which some older guides mention, was discontinued on August 9, 2026.

Can Grok post to X for me automatically?

No, not through its own X connector. The Grok Bot X connector that launched on August 29, 2026 is read-only: it searches posts, reads your timeline and checks mentions. Publishing is not on the list. X owns Grok and Grok still cannot post to X without a custom MCP connector added at grok.com/connectors.

Can Perplexity post to X?

Perplexity Comet cannot. It summarises and drafts, and you copy the text out yourself. Perplexity Computer, launched February 25, 2026 at 200 dollars per month, can post once you add a custom MCP connector. Custom connectors reached Pro and Max in March 2026, and the Mac app runs local stdio servers.

Do I need an X developer account for my AI agent to tweet?

Not if you use a posting layer that holds the X connection for you. You do need one if your agent calls the X API directly, and since April 2026 new developers cannot buy the old Basic or Pro tiers at all. The remaining options are metered pay-per-use at $0.015 per post, $0.20 if the post contains a link, or an Enterprise contract. More detail: do I need an X developer account?

Get Started

Whichever agent you picked from the table, the path is the same five steps.

1. Sign up for OpenTweet (7-day free trial, no credit card)
2. Connect your X account through the dashboard. OpenTweet holds the OAuth connection, so you never apply for a developer account
3. Generate an API key. It starts with ot_
4. Add the MCP server to your agent, or drop a SKILL.md into its skills directory
5. Ask your agent to schedule its first post, then review the queue before it goes out

Building on top of it rather than chatting with it? The REST base is https://opentweet.io/api/v1 and you create a post with POST /api/v1/posts. See the agent docs.

---

The OpenTweet MCP server is open-source and available on npm. Learn more at opentweet.io/mcp.