Buffer AI Assistant: What It Does, Where Data Goes, and MCP Agents

What Buffer's AI is, and what it is not

Buffer is a publishing tool with AI added at the edges, not an AI tool with publishing attached. The company says so plainly: the AI Assistant is optional, and your posts are never generated or changed by AI unless you open the assistant in the composer or the Create space. The core features, from the queue to the calendar, are not AI-powered. That matters if your brand or your clients are wary of automated content. Nothing happens on its own. Around the assistant sit a few smaller helpers: alt text drafts for images, reply suggestions in the comment inbox, short performance notes in Insights, and a toggle that tells networks a post was made with AI.

The second half of the story is newer and, for some teams, more important. Since June 2026 Buffer has run an official MCP server, so assistants such as Claude, ChatGPT and Perplexity can read your channels, draft and schedule posts, and pull analytics from outside Buffer. A command-line tool with skill files for Claude Code and Codex sits on top. Put together, Buffer's AI is less a writing robot than a set of handles. You decide which ones to pull. This guide covers each one, what it does on each plan, where your text goes, and which jobs should stay with a person.

The AI Assistant in the composer

The AI Assistant launched in February 2023 and needs no setup: you open the composer, click the assistant and start typing. It is included on every plan, Free included, with unlimited use ^[1]. It does need an account, so guest mode lacks it, and it belongs to the current plans only; customers still on legacy plans do not get it. Buffer's help center says it runs on OpenAI's GPT-4.0. Buffer's own pages also warn that AI answers can be inaccurate or misleading and should be reviewed before use. Take that warning literally.

Create, refine, repurpose

The assistant does three kinds of work. To create, you give it a prompt of at least four words and under 10,000 characters, and it returns a draft. To refine, you take a draft and press Rephrase, Shorten, Expand, More Casual or More Formal. Inside the composer it is channel-aware, so a rewrite for Instagram stays under the 2,200-character caption limit, and Buffer pitches tailoring for LinkedIn, Threads, X, Instagram and Bluesky. In the Create space, a Generate Ideas button brainstorms post ideas once you describe your business and your target audience. That is a quick way to fill an empty ideas board.

Repurposing comes with one limit you need to know. The assistant cannot open or read any URL you paste. Hand it a link to your blog post and it never sees the article. To turn a long piece into posts, paste the text itself into the prompt, inside the 10,000-character ceiling. Outside the app, Buffer also runs a free blog-to-social tool that takes up to 20,000 characters and returns a hook-led, a summary and a takeaway-led post. Anything you plan to schedule still goes through the composer.

Where the text you type goes

Buffer's answer is short and specific. The assistant is built on the OpenAI API, and the text you enter into it is shared with OpenAI ^[2]. According to Buffer, nothing else goes along: not your account details, not your media files, not the other parts of your post. So the rule for teams is simple: treat whatever you type into the assistant as sent to a third party. Keep embargoed client news, unreleased numbers and anything confidential out of the prompt box. Under Buffer's Terms you own what the AI feature produces, and using it means accepting OpenAI's terms as well, because the Terms fold them in by reference.

Alt text drafts and the AI-generated label

AI alt text, still in beta

Since June 2026 the composer can draft alt text for you. A Generate with AI button reads the image and suggests an editable description of up to about 300 characters, which you review before posting. The feature is labeled Beta, works in the composer and the new image editor, and is not yet available in Ideas. Buffer itself recommends keeping alt text under 200 characters. Expect to trim. On the web, alt text works on static images for X, Mastodon, LinkedIn, Pinterest, Facebook, Instagram and Bluesky, but not on GIFs, Google Business Profile posts or TikTok. On Instagram it applies only to image posts and carousels.

Telling the networks a post was made with AI

The AI-Generated toggle does one job. At publish time it passes an AI disclosure to the network, which then shows its own native AI label. Buffer lists support for Instagram posts, Reels, carousels and Stories, TikTok video posts, YouTube Shorts and original X posts ^[3]. The same help page, oddly, names only TikTok, YouTube Shorts and X in its step-by-step section. Check Instagram on a real post before you rely on it, especially under client disclosure rules.

The edges are clear. TikTok takes the label only on videos, so switching a post to photos removes the toggle. Reminder posts, which you finish on your phone, cannot carry it at all. Sending the flag through the API or the MCP server on an unsupported post, such as a TikTok photo or an X repost or quote, returns a validation error. There is no filter for flagged posts; they only show a sparkles icon in the Queue and the Calendar. On mobile the toggle is available on iOS and still in beta on Android, so a team that posts from phones should check both.

Reply suggestions in Community and Takeaways in Insights

Community, Buffer's comment inbox, has optional AI that drafts replies in your voice and highlights the comments drawing the most engagement, a comment-insights feature listed on every plan but still rolling out. Here sits the clearest plan difference in Buffer's in-app AI: five AI reply suggestions a week on Free, unlimited on Essentials and Team. Suggestions do not work for X at all, because of limits in X's API, while Google Business Profile reviews with a star rating can get them. For a small brand answering a few comments a day, five a week runs out by Tuesday. A paid plan removes the cap. Which networks feed the inbox is covered in our Buffer Community guide.

Insights, the analytics area that replaced Analyze, adds Takeaways: short, readable observations about performance that suggest a next step. They come on every plan, and since September 2026 they also work in the All Channels view, so you get notes across accounts rather than one channel at a time. A few details limit them. Takeaways ignore the date range you pick. Repost suggestions need posts older than four weeks and skip Bluesky and Mastodon. They are web-only. On paid plans, Insights can also export a single Markdown file, which Buffer positions for reports, documents or AI tools.

Recommended posting times come from general data

When you set a date and time for a post, Buffer can suggest a slot for that specific day. Do not read too much into it. Buffer describes the suggestions as based on general engagement patterns and, only on some plans, on recent performance. Its TikTok page goes further and says there is no automatic, account-specific best-time feature for TikTok. Insights itself offers no best-time recommendations yet. A few Buffer pages describe the suggestions more generously, but the plainer wording is the one to plan around.

The general data is large, at least. In September 2026 Buffer published best-time studies built on 9.6 million Instagram posts and over 4.8 million LinkedIn posts sent through Buffer, and on 7.1 million TikTok posts. That is a fair starting point for a new account with no history. It is not a picture of your audience. A local bakery and a software company keep different hours, and so do their followers. Use the suggested times for the first few weeks, then move your posting slots toward the hours your own Insights show working. That call belongs to a person reading the numbers.

The fine print: providers and credits

Two parts of Buffer's legal hub sit awkwardly next to the marketing. The first is the provider list. The Terms name only OpenAI as the model provider for the AI feature, which they call Buffer Creator. The sub-processor list, with 26 entries, names both Anthropic and OpenAI as AI sub-processors in the USA. Buffer does not say which features use which provider beyond the assistant. If your data rules require you to list every AI vendor that may touch your content, list both.

The second is credits. The pricing page says the AI Assistant comes with unlimited credits on every plan, while the Terms say AI use may be limited by credits allocated by subscription level and number of connected channels. Today's offer is unlimited; the contract leaves room to change it. We read that as a reserved right rather than a hidden meter. Still, if your workflow leans heavily on the assistant, do not assume unlimited will last forever. The rest of the product, plan by plan, is covered in our Buffer review.

Connecting Claude, ChatGPT and other agents through MCP

Buffer's MCP server lives at mcp.buffer.com/mcp, and it is available on every plan. Buffer warns that this is the only official endpoint, so third-party one-click installs deserve a second look. Most assistants sign in with OAuth, so no key changes hands: Claude, ChatGPT and the Codex CLI, Cursor, Grok, Manus and Notion custom agents all connect this way, though ChatGPT on the web needs Developer mode switched on. n8n and Google's Antigravity use a Buffer API key sent as a Bearer header, and Raycast requires Raycast Pro. Any client that can send headers can use a key. You see and revoke connections under Settings, then API.

What an agent can do

An agent connected this way can read your channels, browse the queue and drafts, create, schedule and edit posts, capture ideas and pull post analytics ^[4]. It can also work with post templates and create private or organization-wide ones. A built-in prompt reviews your week of posts, and two generic tools reach the rest of the GraphQL API. Every ID an agent sends must match a 24-character hex pattern, so invented IDs are rejected before they cost a request. Analytics pulled this way refresh once a day and can lag the network by up to a day. Fine for weekly reviews. Not live reporting.

Shared quotas and account-wide reach

Two facts shape every agent setup. First, MCP requests share the same quota as your API keys: 100 requests per 15 minutes on every plan, 250 per 24 hours on Free and Essentials and 500 on Team, and 3,000, 7,500 or 15,000 per 30 days on Free, Essentials and Team. Connecting a second assistant adds no capacity. Second, a connection covers the whole account, with no per-organization scoping, so an agent you connect for one client can see the others. For agencies, that is a real risk, so plan the account structure before anyone connects an assistant.

Buffer is candid about agent behavior. Agents tend to rephrase and retry a failed call instead of waiting, so tell yours to stop on a rate-limit error and report the wait time. Write tools take no idempotency key, so retrying a post whose reply was lost can publish a duplicate. Buffer advises keeping approval prompts on, because creating a post can publish at once and deleting one cannot be undone. One safeguard helps: posts from users set to Needs Approval always land as drafts, even through MCP. The full developer picture is in our Buffer API guide.

The CLI and skill files for Claude Code and Codex

For developers and terminal-minded marketers, Buffer ships a command-line tool as an npm package that needs Node.js 18 or later. Its commands are generated from the public GraphQL schema and return structured JSON, which suits an agent reading the output. Two commands install Markdown skill files, one for Claude Code and one for Codex, so the coding agent knows how to drive Buffer without you explaining the API. Every write command supports a dry run that checks the payload locally and prints it without calling Buffer. Use it before any bulk job.

The CLI also guards your quota. It warns when you are down to 10% of your remaining requests and exits with a distinct code when the API says you have hit the limit. Your key can be stored only in the global config and is stripped from repository config, which keeps it out of shared code. Buffer's agent setup guide adds sensible house rules: never paste the key into chat, read it from an environment variable, make only read-only requests during setup, and get explicit approval before every change.

The playbook: what to hand to AI and what to keep human

Buffer's AI is useful for volume and risky for judgment, so split the work that way. The assistant can be wrong, the reply helpers cannot read a situation, and an agent with approvals off can publish or delete before anyone looks. Everything below follows from those three facts. Treat the first list as drafting help and the second as decisions that need a named person.

Hand to AI:

• First drafts from a short brief, then Shorten or Rephrase per network.
• Pasted-in repurposing of a blog post or newsletter.
• Alt text drafts for every image, trimmed and checked by eye.
• Reply drafts for routine comments, such as thanks or opening hours.
• A weekly review of what went out, through Takeaways or an agent.

Keep human:

• The final read of every post before it publishes.
• Replies to complaints, crises and anything legal or medical.
• Anything confidential, because the assistant's text goes to OpenAI.
• Posting times once your own data exists.
• Deleting posts and approving client content.

A good rhythm for a small team looks like this. On Monday, an agent drafts the week from your notes and parks the posts as drafts. A person edits, adds alt text from the AI draft, flags AI-made media with the toggle and approves. During the week, reply suggestions handle the routine comments while a person takes the hard ones. On Friday, Takeaways or the agent summarize what worked, and a person decides what changes next week.

The bottom line

Buffer's AI is modest and honest about it. The assistant is unlimited on every current plan, works only when you open it, and sends what you type to OpenAI and, Buffer says, nothing else. Alt text, reply suggestions, Takeaways and the AI label are small helpers that save minutes rather than replace anyone. The bigger shift is the MCP server and CLI, which let Claude, ChatGPT and other agents run your queue from outside Buffer. Use them with approvals on, keep the shared quota and account-wide scope in mind, and leave judgment, crises and final reads with a person.