State of AI Agents in Social Media 2026: MCP and API Data From 35,000+ Posts

How many social media posts are created by AI agents and software in 2026? On PostEverywhere, most of them. Between 1 August and 30 September 2026, 79% of posts were created through our API or MCP server rather than by a person clicking in the web app. Most of that volume comes from code and automations; about 15% of API post requests came directly from an AI assistant through MCP. And the assistant side is the fastest-growing part: the number of organisations posting through an AI assistant nearly tripled between July and September.

This is the first State of AI Agents in Social Media report. Every number below comes from PostEverywhere's own first-party usage data: 35,000+ posts from 330+ organisations, published to all top platforms.

It is data only a scheduler with real traffic through the Model Context Protocol can publish, and we have not seen it published elsewhere.

We wrote it because the question "are AI agents actually posting to social media yet, or is it still a demo?" deserves a measured answer rather than another prediction.

Key findings on AI agents in social media#

• 79% of posts were created by software, not a person in an app. 20% came from the web app and under 1% from our iOS app.
• MCP adoption nearly tripled in two months. Organisations creating posts through an AI assistant grew 2.8x from July to September 2026, and MCP's share of all organisations posting through our API rose from 47% to 62%.
• Two thirds of teams that post programmatically use MCP. 66% of organisations that sent post-creation requests to our API used an AI assistant for at least some of them, and 35% used nothing but an assistant.
• Claude dominates. 83% of AI assistant connections to PostEverywhere were Claude. ChatGPT was 8%.
• 64% of people who connect an assistant go on to publish through it. Connecting is not a novelty click.
• Agents and scripts post to one platform at a time. A post created through the API or MCP targets 1.5 platforms on average. A post made in the app targets 3.5.
• Software and people favour different platforms. API and MCP posts lead with Facebook and Instagram. People in the app lead with LinkedIn.
• Agents schedule and revise. The typical API organisation schedules 100% of its posts, and assistants send about 8 edit requests for every 10 post-creation requests.

The rest of this report explains each finding, what we think it means, and how we measured it. The full methodology and its limits are at the end.

Finding 1: most posts on PostEverywhere are now created by software#

We split every post created on PostEverywhere by the surface it came from: the web app, the iOS app, or the API. The API surface includes both our MCP server, which is how Claude, ChatGPT and other assistants post, and direct calls to our social media API from code people write.

The share stayed between 68% and 84% in every full week of the period.

It is not one heavy customer skewing the number. The ten largest API organisations account for 45% of API posts, so more than half comes from the long tail. Looking at organisations rather than posts:

• 50% of active organisations used only the web or iOS app
• 35% used only the API or MCP
• 15% used both

So half of our active organisations now create posts through software, and those that do post far more. The typical (median) API organisation created about 4.7 times as many posts as the typical app organisation over the same two months.

What we think it means: social media publishing is splitting into two jobs. People still use a visual composer to plan and review. The volume, the daily and weekly repeat work, is moving to agents and scripts.

If your content operation publishes more than a handful of posts a week, the question is no longer whether to automate but which surface to automate through. Our API vs MCP guide walks through that decision.

Want your agent to post for you? Connect Claude, ChatGPT or any MCP client to PostEverywhere in about two minutes and publish to all top platforms from a chat. See how AI agents connect to PostEverywhere.

Finding 2: MCP adoption nearly tripled in two months#

MCP is the open standard Anthropic introduced in November 2024 for connecting AI assistants to tools. In 2026 it became the main way assistants take actions in PostEverywhere.

We tracked how many organisations created at least one post through our MCP server each month. Our tracking of MCP requests began on 13 June 2026, so we use July, the first full month, as the base. Indexed to July 2026 = 100:

Organisations posting through an assistant grew 2.8x in two months, and posts created through MCP grew 3x. The last column matters most: MCP is not just growing along with PostEverywhere. It is taking a bigger share of the organisations that post programmatically each month.

New assistant connections tell the same story: people connecting an AI assistant to PostEverywhere for the first time rose about 5x from July to September.

The wider API picture is growing too. The number of organisations posting through the API in a given week rose about 2.5x in seven weeks, from the first full week of August to the third week of September, while the number of all active organisations rose about 1.5x.

What we think it means: MCP crossed from developer curiosity to a normal way of working in the space of one summer. For a growing share of our users, the question has changed from "does it have an API?" to "can my assistant use it?".

Our roundup of the best social media MCP servers compares what the options can actually do.

Finding 3: two thirds of programmatic posters use MCP, but code still drives the volume#

The API surface covers several very different ways of posting. We classified every post-creation request by the client that sent it.

By organisation, MCP is the most common way to post programmatically. Of organisations that sent post-creation requests to our API:

• 66% used MCP (an AI assistant) for at least some of them
• 35% used MCP only, with no other client at all
• 31% mixed MCP with other clients
• 34% never used MCP, only their own code or an automation tool

By volume, code still dominates. About 74% of API post-creation requests came from clients we could not attribute to MCP, our SDK or a named automation platform, which is mostly code customers wrote themselves (Python, Node.js, curl and similar).

About 15% came from MCP, and about 10% from our SDK, CLI and automation platforms such as n8n and Make.

What we think it means: these are two different ways of working. MCP is how most teams start, because it needs no code: you connect an assistant and ask it to post.

Custom code is how the heaviest publishing runs, where posts flow in from a team's own systems every day. If you are building that second kind, our guide to building a social media agent with the API is the place to start.

Finding 4: Claude is the dominant AI agent for social media#

When someone connects an AI assistant to PostEverywhere, we record which client they connected. Across every assistant connection from 13 June to 30 September 2026:

Counting Claude Code with Claude, Anthropic's clients account for about 86% of connections.

Connecting is not a one-off experiment. 64% of people who connected an assistant went on to publish a post through it.

What we think it means: this is our reading, not something the data proves. Claude's connector system makes adding a remote MCP server a few clicks, and the same connector works across claude.ai, Desktop, mobile and Claude Code.

(For the other connectors marketers add alongside a publishing one, see our top 10 Claude connectors.) ChatGPT supports MCP too, as OpenAI's MCP documentation describes, but in our data far fewer people have gone through with it so far.

If you use Claude, our Claude connection guide takes about two minutes. ChatGPT users can follow the ChatGPT guide.

We expect the gap to narrow as other assistants make connectors easier to add. It is the number in this report we will watch most closely in the next edition.

Using Claude already? Add PostEverywhere as a connector and ask Claude to draft, schedule and publish your week of posts. Plans start at $9/month with a 7-day trial. Compare plans.

Finding 5: agents and scripts post to one platform at a time#

This was the finding we least expected. When people create a post in our web app, they usually send it to several platforms at once. When a post arrives through the API or MCP, it usually targets one.

What we think it means: our reading is that agents and scripts often write a separate post for each platform, because they can: a LinkedIn version, an Instagram version, an X (Twitter) version, each adapted to that platform.

A person, by contrast, tends to write one caption and tick five platforms. The result looks less like one message broadcast everywhere and more like a set of tailored posts that go out together.

That matters for anyone choosing tools. Software posting this way needs a tool that accepts per-platform content, per-platform media and per-platform settings in a single account, which is how our cross-posting works under the hood.

It also explains why the manual version of this job, adapting one post for every network, is where people lose most of their time. Our comparison of cross-posting tools covers the manual side.

Finding 6: software and people favour different platforms#

We counted every platform each post was sent to and compared the mix by surface.

The difference is LinkedIn. It is the top platform for people posting in the app, at nearly 30% of their platform targets, and only fifth for posts arriving through the API and MCP. Software leans towards Facebook, Instagram and X (Twitter).

What we think it means: our reading is that people still want a hand on the wheel for LinkedIn, where posts are personal and tied to someone's professional reputation. The high-frequency, brand-led networks, where a business publishes daily across Facebook and Instagram, are where our users most often hand the work to software.

If you post to LinkedIn by hand today, that is normal.

Finding 7: agents schedule ahead and post less video#

Two more behaviours separate API and MCP posting from app posting.

They schedule. The typical organisation posting through the API scheduled 100% of its posts for later rather than publishing immediately. The typical app organisation scheduled 79%. Across all posts the gap is smaller, 70% versus 75%, likely because a few high-volume API organisations publish immediately from their own systems.

They post less video. About a third of API and MCP posts included a video (34%), compared with 60% of app posts. Nearly all posts in both groups carried media of some kind (90% versus 92%), so software is posting images rather than going text-only.

Caption length, by contrast, was almost identical: the median post in both groups was about 290 to 300 characters.

What we think it means: agents and scripts are good at the planning and packaging work (writing a caption, attaching an image, picking a time) and are used less where the creative asset is the hard part. Video still mostly starts with a person.

That is also why an assistant that can generate an image as part of the same conversation matters more to agent workflows than to app workflows.

Finding 8: what AI agents actually do inside a social media tool#

Finally, we looked at which actions AI assistants take through our MCP server. Share of organisations using MCP that performed each action at least once (Aug to Sep 2026):

Two patterns stand out.

• Agents revise. Assistants sent about 8 edit requests for every 10 post-creation requests. People ask for a draft, read it, and ask for changes before it goes out. The agent is a collaborator in the loop, not a fire-and-forget bot.
• A third of agent users check results. One in three MCP organisations checked whether a post actually published, and a similar share read analytics. Agents are starting to close the loop, not just publish.

What we think it means: the review step is the whole game. The safest agent setups keep a person approving what goes out, which is the approach we describe in how to give AI agents safe access to social media.

Build this workflow for your team. PostEverywhere gives your assistant drafting, editing, scheduling and results checking through one connection, with every post visible in your calendar before it publishes. Start with the AI agents setup.

What this means for social media teams#

Three practical conclusions from the data.

1. Automate the repeat work first. Our users hand software the high-frequency, brand-led posting (Facebook, Instagram, X (Twitter)) and keep the personal channels (LinkedIn) closer to hand. That is a sensible order to adopt in.
2. Let the agent tailor, not broadcast. Posts from software usually target one platform at a time. Give your agent the platform rules (caption lengths, media formats) and let it write a version for each network, which people rarely have time to do.
3. Keep a person in the loop. Nearly half of MCP organisations edit posts after the first draft. Schedule rather than publish instantly, and review the calendar before posts go out.

If you are starting from zero, our list of AI agents for social media compares the options by setup effort; the Claude Code MCP guide, Cursor guide and Codex setup guide cover developer setups; and self-hosted agents can connect through OpenClaw.

Our skills and prompts give any of them a ready-made playbook for the repeat work in point 1. Two September 2026 launches widen the field: Meta Muse, Meta's personal agent (rolling out in the US), and OpenAI Dots, always-on agents in ChatGPT. Both connect to PostEverywhere as a custom connector.

What this means for social media tools#

AI assistants are becoming a front door to software. In our data, half of active organisations already create posts through the API or MCP, and a third created every post that way, never through our apps.

For social media tools, that changes what "good" means. A clean dashboard still matters for the 50% of organisations that only use the app. But for the other half, what matters is how well the tool works when an assistant or script is driving: clear per-platform rules, safe scheduling, edits, results the agent can read back, and a developer surface that code can call directly.

Most research in our category measures what people post and how audiences respond. Very little measures how posts get made. That is the gap this report starts to fill.

Methodology and limitations#

Data source. PostEverywhere's first-party product usage events, recorded when a post is created, scheduled or published, when an API request is made, and when an AI assistant is connected. We used usage metadata only (surface, platforms targeted, media type, caption length, action type).

We did not use any content of posts, and we did not use engagement or audience data from the social platforms.

Period. 1 August to 30 September 2026 for all post-level findings, the period in which every post carries a reliable surface label. Growth figures in Finding 2 use monthly data from July to September 2026 (MCP request tracking began on 13 June 2026).

Assistant connection figures in Finding 4 cover every connection from 13 June to 30 September 2026.

Sample. 35,000+ posts from 330+ organisations. We report percentages, medians and indexed growth rather than absolute counts.

Exclusions. PostEverywhere's own brand and test accounts and test connections, and a small group of accounts we identified as abusing trials.

Definitions.

• A post is one publish request. A single post can be sent to several platforms, and platform shares in Finding 6 count each platform a post targets.
• API and MCP covers every post created through our API, including AI assistants connected through our MCP server, our SDK and CLI, automation platforms such as n8n and Make, and code customers wrote themselves.
• MCP means requests sent by our MCP server on behalf of an AI assistant, using the open MCP specification. The assistant (Claude, ChatGPT and so on) is recorded when the person connects it.
• Median organisation figures for scheduling and multi-platform behaviour (Findings 5 and 7) only include organisations with at least five posts on that surface. The 4.7x post-volume comparison in Finding 1 includes all active organisations.

Limitations.

• This is PostEverywhere's user base, not the whole market. We are positioned as a scheduling tool for AI agents, so our users adopt agents faster than the average social media team. Read these numbers as where early adopters are heading, not where everyone is today.
• Findings 5 to 7 cover everything created through the API, which includes code customers wrote as well as AI assistants.
• Assistant shares (Finding 4) are shares of connections, and ChatGPT and the smaller clients are small groups.
• API post counts include automated systems at a small number of high-volume organisations. We report organisation-level medians alongside post-level shares so no single customer drives a conclusion.
• We cannot see what an assistant was asked to do, only which actions it took.

We plan to update this report as the data matures. If you are a journalist or researcher and want to cite or discuss the data, you are welcome to quote any figure with a link to this page.

FAQ#

How many social media posts are created by AI agents?#

On PostEverywhere, 79% of posts created between August and September 2026 came through our API or MCP server rather than the web app. Most of that volume is code and automations; about 15% of API post requests came directly from an AI assistant through MCP.

Two thirds of the organisations posting programmatically used an assistant for at least some of their posts.

How fast is MCP adoption growing for social media?#

The number of organisations creating posts through our MCP server grew 2.8x from July to September 2026, posts created through MCP grew 3x, and MCP's share of organisations posting through our API rose from 47% to 62%.

New assistant connections grew about 5x over the same two months.

Which AI assistant is used most for social media posting?#

Claude, by a wide margin. 83% of AI assistant connections to PostEverywhere were Claude, 8% were ChatGPT, and 10% were other clients such as Claude Code, Codex, Grok, Cursor and Perplexity.

Do AI agents cross-post to several platforms at once?#

Usually not. Posts created through the API and MCP targeted 1.5 platforms on average, compared with 3.5 for posts made in the web app. We believe agents and scripts often write a separate, tailored post for each platform instead of sending one post everywhere.

Which social media platforms do AI agents post to most?#

Facebook (22.5% of platforms targeted by API and MCP posts) and Instagram (20.2%), followed by TikTok and X (Twitter). People posting in the web app lead with LinkedIn, which is nearly 30% of their platform targets but only about 11% of API and MCP targets.

Where does the data in this report come from?#

From PostEverywhere's own product usage data for 1 August to 30 September 2026, covering 35,000+ posts from 330+ organisations. We used usage metadata only, excluded our own accounts and accounts abusing trials, and report percentages rather than absolute counts. The full methodology is above.