// HOW-TO · PUBLISHING

How to publish social posts directly from an AI assistant (2026)

Publish or schedule social posts directly from an AI assistant like Claude or ChatGPT using an MCP server: connect accounts, add the connector, post in chat.

Last verified · 2026-10-04 · by Moe Ameen

As of 2026 you can skip the dashboard entirely for a single post: draft it inside an AI assistant and have the assistant publish or schedule it straight to your connected accounts. The mechanism is the Model Context Protocol (MCP) — an open standard that lets an assistant like Claude or ChatGPT call external tools. You connect it to a social-media MCP server that already holds your account credentials, and the assistant gains a small menu of actions: publish a post now, schedule one for later, list your connected accounts, and check a post's status.

This walkthrough gets you from zero to a published test post, then a scheduled one, and covers how to manage and revoke the access afterward. It's the fast path for one-off posts and replies. For the bigger picture of what this workflow is and where it breaks down, read the companion guide on [social media MCP](/guides/social-media-mcp); to run an assistant as an ongoing manager rather than a publish button, see [set up an AI assistant as your social media manager](/how-to/set-up-an-ai-assistant-as-your-social-media-manager).

The steps

Pick a social-media MCP server. The assistant is only the interface — the server decides which platforms you can reach. Choose one whose platform coverage matches where you post and that publishes through each platform's official API rather than browser automation. Several exist (hosted and self-run) covering overlapping lists across the major networks; pick on coverage, on who operates it, and on whether you can revoke access cleanly.
Connect your accounts to the server. Before the assistant can post anything, the server needs permission to post on your behalf. You do this once, inside the server's own setup, almost always via OAuth: you log into each platform and grant the server posting access. This is the step you're actually trusting — the server holds the keys; the assistant just asks it to act. Connect only the accounts you want the assistant to touch.
Add the MCP connector to your assistant. In Claude this is a custom connector (remote MCP) added in settings, available on every plan from Free up. In ChatGPT it's a connector added through Settings → Connectors → Advanced → Developer mode — but check your plan first: OpenAI currently limits full read-and-write MCP, the kind that can actually publish a post, to Business, Enterprise, and Edu workspaces; an individual Plus or Pro account's developer mode is read/fetch only, so it can list your accounts and check status but can't call a publish or schedule tool. Paste the server's MCP URL and any token it gives you, and confirm the connection grants write access before you count on it to post.
Confirm the assistant can see your accounts. Before posting anything live, ask the assistant to list your connected accounts — this calls the server's list-accounts tool and confirms the wiring end to end. You should see exactly the accounts you authorized. If an account is missing, the problem is in the OAuth connection from step 2, not the assistant; fix it on the server side before continuing.
Publish a test post. Start with something low-stakes on one account. Prompt plainly: "Publish this to my LinkedIn: [text]." The assistant translates that into the server's publish tool and the post goes live. Watch the assistant's confirmation and then check the platform itself — don't trust the chat's "done" without looking. Attach an image in the prompt if you're testing media, since media handling varies by server.
Schedule a post instead of publishing now. To queue rather than ship, give a time: "Schedule this to TikTok and Threads for tomorrow at 9am." That invokes the schedule tool rather than publish. Scheduling is also your only built-in safety margin here — a scheduled post can usually be cancelled before its time, whereas an immediate publish is live the instant the tool returns. Confirm the scheduled time landed in the timezone you meant.
Verify status and manage access. Ask the assistant to check the status of what it submitted — the check-status tool reports whether each job actually succeeded, which matters on multi-platform posts where one network can fail while others succeed. When you're done, know how to revoke: you can remove the connector from the assistant and, separately, revoke the server's OAuth access from each platform's connected-apps settings. Do both if you're cutting a server off entirely.

Common gotchas

No review queue. "Prompt, confirm, live" has no draft state — when the assistant publishes, it publishes. Treat every live-publish prompt as irreversible and proofread before you send it, not after.
No brand guard. Nothing between your prompt and the post checks voice, banned words, disclosures, or compliance. The model will happily ship something off-brand because nothing in the loop is responsible for catching it.
The assistant can misread the prompt. A vague "post this everywhere" can hit the wrong accounts or the wrong platforms. Name the exact accounts, and confirm with the list-accounts tool first.
OAuth is broad access. Connecting a server grants it ongoing posting rights to your accounts, not a one-time pass. Audit which server holds what, and revoke cleanly when you stop using one.
Coverage is the server's, not the assistant's. If a platform won't post, it's almost always that the server doesn't support it — check the server's platform list before blaming the chatbot.
It doesn't scale to a content program. A chat is great for one post and tedious for a week of multi-format content across eight platforms — that's a pipeline job, not a conversation.

Legal note

Publishing through an MCP server means granting a third party ongoing OAuth access to post to your accounts — review the server operator and scope before connecting, and revoke access from each platform's connected-apps settings when you stop using it. Posting via official platform APIs is sanctioned; servers that rely on browser automation to reach unsupported platforms may conflict with those platforms' terms of service. You remain responsible for every post an assistant publishes on your behalf, including disclosure and compliance requirements in your industry.

Where Kompozy fits

Notice what the MCP server you just wired up actually does: it performs the publish and nothing else. It does not write the post, it does not know your brand, and it does not care whether you're shipping one text update or the tenth off-voice draft in a row — it's a pipe, and it faithfully pushes whatever you hand it. That's the right design for a one-off. It's also exactly why the moment you want this same "it posts for me" ease applied to recurring, on-brand content, you've outgrown the chat — because the hard part was never the publish, it was everything that should happen before it.

That before-the-publish half is what [Kompozy](/) is. Where the MCP server takes whatever text you typed, Kompozy generates the content against a [Persona Brief](/glossary/persona-brief) that governs voice and runs a banned-word and prohibited-topic filter, so what ships is on-brand by construction rather than by you proofreading each prompt. And where a chat does one text post, Kompozy turns a single idea into net-new formats the MCP publish can't make on its own — a captioned [Persona Short](/glossary/persona-shorts) or avatar video, brand-exact carousels via [HyperFrames](/glossary/hyperframes), images, a blog, a newsletter — across [18 output formats](/glossary/output-buckets) in one pass.

The publishing destination is the same kind of official-API plumbing the MCP server uses — Kompozy fans output to the eight social platforms plus blog and email — but driven by [Autopilot](/glossary/autopilot) on a set cadence and gated by a per-post review step, the two things the chat removes. So the clean split is this: keep the MCP-connected assistant for genuine one-offs and timely replies, and when the job becomes a recurring program, let Kompozy be the brain and the cadence the dumb pipe was never meant to supply.

Frequently asked questions

Can Claude and ChatGPT really publish to social media?

Yes, through MCP custom connectors, with one plan caveat on the ChatGPT side. Neither chatbot posts on its own — you connect it to a social-media MCP server that holds your account credentials, and the assistant calls that server's publish and schedule tools. Claude supports write-capable remote MCP connectors on every plan, Free included. ChatGPT's developer mode currently restricts full read-and-write MCP to Business, Enterprise, and Edu workspaces; an individual Plus or Pro account gets read/fetch-only access, so it can check your connected accounts but can't yet fire the publish tool itself. The platforms you can reach also depend on the server you connect, not just the assistant.

Do I need to give the AI assistant my social passwords?

No. The assistant never sees your passwords. You connect your accounts to the MCP server once, almost always via OAuth, and the server holds that access. The assistant only asks the server to perform named actions like publish or schedule — it never talks to the platforms directly or stores your credentials.

What can go wrong publishing straight from a chat?

The three structural gaps are no review queue (a published post is immediately live, with no draft state to catch errors), no brand or compliance guard between your prompt and the post, and a conversational interface that's poor at managing more than one post at a time. For a one-off it's fine; for recurring brand content the missing review gate is the real exposure.

Is publishing from an AI assistant good for a brand's whole social presence?

No — it's the wrong shape of tool for that. Direct-from-chat publishing is a precision tool for a single post, reply, or experiment. A sustained brand presence needs recurring multi-format generation, brand governance, scheduling across platforms, and a human review gate before anything ships — a production pipeline, not a chat window. Use the assistant for one-offs and an engine for the ongoing program.