AI Social Media Manager: How to Run One with an Agent

TLDR: An AI social media manager in 2026 is an agent such as Claude connected to a publishing API through MCP, so it can plan, draft, schedule, read analytics, and triage comments while a person reviews every post. Start read-only and limit profile access. Creating drafts needs write access, which can also publish, so the approval rule rests on your instructions.

Search for the phrase and the results are lists of scheduling tools with a writing button: paste a topic, get captions, click schedule. That is one meaning of an AI social media manager, and it serves plenty of teams. The other meaning, and the subject of this post, is an agent you already use, Claude, Cursor, Codex, or OpenClaw, connected to a publishing API so it can read your analytics, draft posts for each platform, fill a queue, and sort comments, with you approving before anything goes live.

We build the publishing layer those agents call, so we see what works and where it goes wrong. This post is for marketers and founders who run their own accounts and developers connecting an agent for someone else. The setup commands are at the end.

What is an AI social media manager?

A social media manager plans the week, drafts and schedules posts, publishes, reviews results, answers comments, and adjusts the next plan. AI software can handle parts of that workflow, and how much depends on which of two shapes you pick.

The first shape is an AI writing feature inside a scheduling tool. It works in the composer: you press a button, and a model writes or rewrites the caption. You still operate the calendar, the publishing controls, and the analytics screens. It is a reasonable choice: nothing to configure, no key to protect, and the person at the keyboard makes every decision. If your week is ten posts and the hard part is the writing, it may be all you need.

The second shape is an agent with tools. The model runs inside a client you already use (claude.ai, Claude Desktop, Claude Code, Cursor, Codex, or OpenClaw), connected to a publishing API. It can call listPosts to see what went out, getAnalytics to see how each post did, getBestTimeToPost to suggest a slot, createPost to write a draft, and listInboxPostComments to read replies.

The second shape is what the rest of this post means by an AI social media manager. It is more capable, and it needs controls because it can act on your accounts.

How does an agent manage social media through an API?

Your AI client calls PostZen’s tools, and PostZen communicates with the social platforms:

You: prompts in, approvals out
        |
        v
AI client: claude.ai, Claude Desktop, Claude Code, Cursor, Codex, OpenClaw
        |
        |  MCP: tool calls over HTTPS
        v
PostZen MCP server: https://mcp.postzen.dev/mcp
        |
        |  the same operations as the REST API
        v
PostZen API: posts, queues, media, inbox, analytics, webhooks
        |
        v
Instagram, X, LinkedIn, Facebook, Threads, YouTube, TikTok,
Pinterest, Bluesky, Telegram

Approval step: the agent creates posts with isDraft: true.
A person reviews each draft in the dashboard and schedules or publishes it.

MCP is the Model Context Protocol, an open protocol that lets an AI assistant discover and call tools provided by a server. The server describes each tool with a name, a description, and an input schema specifying the data it accepts. The client gives those descriptions to the model, which decides when to call the tools.

Our MCP server is hosted at https://mcp.postzen.dev/mcp, with nothing to install. As of October 19, 2026, it exposes 69 tools: 68 generated from our OpenAPI spec, so every REST endpoint has a matching tool, plus verifyApiKey. They cover:

• profiles, accounts, and connect links
• posts: create, read, update, delete
• queue slots and queue previews
• media upload links
• comments on Instagram, Facebook Pages, and Threads, plus Instagram DMs
• comment-to-DM automations and contacts
• analytics, webhooks, Pinterest boards, and API keys

Analytics tools include post analytics, daily metrics, best time to post, follower stats, and post timelines. The agent can publish to all 10 platforms through the same tools, using the scheduler described in our social media scheduling API guide. Inbox and analytics coverage varies by platform, as detailed below.

You provide the instructions and review the drafts. A post created with isDraft: true appears in your dashboard, where you can edit, schedule, or delete it. It remains a draft until someone or an agent with write access schedules or publishes it.

Which AI agent should you use?

We recommend starting with the client you already use. Every client below connects to remote MCP servers, so you can choose based on where you work.

• Claude in claude.ai or Claude Desktop lets marketers add the connector in Settings and use it through chat. Claude Code provides a terminal interface for developers. Setup for both is on our Claude page.
• Cursor reads a JSON block in .cursor/mcp.json, so the connection configuration can live in the repository. Our Cursor page has the block.
• Codex: setup is on the Codex page.
• OpenClaw: setup is on the agents overview page.

Our agents overview lists all of them, plus GrokBot and Hermes. Our setup pages for Codex, OpenClaw, GrokBot, and Hermes follow the setup patterns those clients publish. We do not rank the clients and have not measured which one writes better posts. Our opinion is that the instruction file and the guardrails matter more than the model.

What guardrails does an AI social media manager need?

An agent with a Read & Write key can publish to every account within that key’s scope. Set these controls before giving it write access.

Start read-only

When you approve the connection at app.postzen.dev/connect/mcp, the consent screen offers Read & Write, the default, or Read Only. Read Only permits fetching posts, analytics, comments, and queue previews. Run the first week that way: the agent learns your accounts, you learn how it reasons, and the worst outcome of a bad prompt is a wrong summary. Scripted setups offer the same choice when you create a pzn_live_... key.

Scope it to the profiles it should touch

The same screen offers access to every profile, the default, or a selected set. If the agent runs one brand, give it access only to that brand’s profile. A key limited to selected profiles cannot list, draft, or publish to the others, whatever the prompt says.

Use drafts as the approval step

Tell the agent to create every post with isDraft: true and never with publishNow. Drafts appear in the dashboard for a person to review, edit, and schedule. This is the approval workflow we use ourselves.

PostZen has no draft-only permission level. A Read & Write key can publish, and a Read Only key cannot create drafts. The requirement to get approval before publishing therefore depends on your instructions and the agent following them. Put the rule in the instruction file described below, repeat it in any prompt that creates posts, and review drafts before scheduling them.

Read the tool annotations

Every PostZen MCP tool carries annotations, including a read-only hint on the 30 generated tools that only fetch and a destructive hint on the 20 that delete or overwrite, such as deletePost and deleteWebhook. Clients can use these hints to decide whether to ask before running a tool. Check your client’s confirmation settings. The hints themselves do not enforce confirmation.

Keep keys revocable

Each OAuth approval creates its own API key, named MCP: followed by the client’s name, on the dashboard’s API Keys page. Revoking that key removes that client’s access. A pending consent request expires after 10 minutes if nobody approves it.

Make creates idempotent

An idempotent request can be retried without creating a duplicate. This matters because an agent may retry a tool call after a timeout even if the first call created the post. Our create endpoint accepts an x-request-id header: send the same ID on a retry, and the API returns the original post instead of making a second one. Set it on every create in your own scripts. Through MCP, instruct the agent to check listPosts after a failed create before trying again.

Pin the brand rules

Write short instructions the agent reads at the start of every session. Include the brand voice, claims it must never make, links it may use, hashtag rules, platform requirements such as X’s character limit and Instagram’s image requirement, and the standing rule that every post is a draft. Claude Desktop and claude.ai accept project instructions, Claude Code reads a CLAUDE.md, and Cursor reads its rules files. Without it the agent guesses, and it guesses like a generic marketer.

Keep customer replies human

The inbox tools can read and reply to comments and Instagram DMs. Let the agent draft replies. Do not let it send them unsupervised, and never on refunds, outages, health, legal matters, or an angry customer. The prompt pack below has it hand you suggested replies to approve. Comment-to-DM automations use a fixed message you wrote, so they do not involve a model composing a new reply each time.

A week of prompts for an AI social media manager

These prompts specify tasks, profiles, time windows, and limits on what the agent should do. Replace Acme with your profile name and adjust the platforms to match your connected accounts.

MONDAY: plan from last week's numbers
List every post published between last Monday and Sunday on the Acme
profile. Get analytics for each one and rank them by engagement. Tell me
the three themes that did best and the two that did worst, with the
numbers from the tools, not estimates. Do not create any posts.

TUESDAY: repurpose a blog post
Read this article: [URL]. Write one post per platform for Instagram, X
(under 280 characters, no hashtags), Threads, and LinkedIn, in the voice
from the brand file. Create each one on the Acme profile as a draft with
isDraft true and give me the draft ids. For Instagram, attach the
article's image if it has one; otherwise give me the caption and an
image brief instead of creating that draft. Do not schedule or publish.

WEDNESDAY: fill the queue
Preview the Acme profile's queue for the next 7 days and list the empty
slots by day and platform. For each empty slot, create a draft (isDraft
true) that fits that platform, using the themes that ranked best on
Monday. For Instagram, TikTok, or Pinterest slots, attach an image from
the original post you draw on if there is one; otherwise give me the
caption and an image brief instead of a draft. Report each draft id with
the slot it is meant for, so I can schedule it from the dashboard. Leave
slots empty rather than guessing.

THURSDAY: triage comments
List comments from the last 48 hours on Acme's Instagram,
Facebook Page, and Threads posts. Group them: questions, praise,
complaints, spam. For every question and complaint, write a suggested
reply in the brand voice and show it to me with the original comment.
Do not post any reply. I will approve them one at a time.

FRIDAY: read the week
Get daily metrics and follower stats for Acme's Instagram and X accounts
for the last 14 days and compare this week to last week. Give me a
six-line summary I can paste into Slack: what moved, the likely cause
from the posts that went out, and one change for next week. Numbers
from the tools only; say "no data" where a platform returns none.

SATURDAY: recycle evergreen posts
List Acme's posts from 60 to 180 days ago and pick the five with the
highest engagement that are still accurate today. Rewrite each one for
a platform it has not run on before. Create them as drafts (isDraft
true) and note which original each one came from. For Instagram,
TikTok, or Pinterest, attach the original post's image if there is one;
otherwise give me the caption and an image brief instead of a draft.
Do not publish.

SUNDAY: next week's calendar
Call best time to post for each Acme account. Draft next week's
calendar as a table: one post per weekday per platform, with topic,
platform, suggested time from the tool, and a one-line hook. Create
nothing yet. I will pick the rows to turn into drafts.

Run Monday first and keep its answer available for later planning, particularly Wednesday’s prompt. Keep the “do not” sentence at the end of each prompt even after the agent has behaved for a month; the key does not enforce it, so it is the only thing standing between a Read & Write key and a published post. If Tuesday’s drafts sound generic, review the brand file. Queue prompts also need configured slots; our queue scheduling page shows how to set them up.

What can an AI social media manager not do well?

These are the limits we see from running the publishing side, as of October 19, 2026.

Taste and judgment. A fluent post can still be wrong for the moment or the brand. The agent may miss that a joke lands badly this week, that a competitor just had an outage, or that a founder would never phrase something that way. Sensitive topics need a person’s judgment.

Platform state it cannot see. The publishing API can report a failed post without explaining the cause in terms the agent can act on. Instagram can drop a caption or hide a reel on an account it has flagged, even when the caption was sent; we have seen it in a support case. Accounts can also be restricted and tokens can expire. The agent’s view is limited to what the tools report.

Visuals. In our experience, agents describe images more reliably than they design them. Instagram, Pinterest, and TikTok need media on every post, drafts included, so the prompts ask for an image brief whenever there is no image to attach.

Analytics gaps. Our analytics tools are live for Instagram, X, Facebook, Threads, YouTube, TikTok, Pinterest, and Bluesky. They are not available for LinkedIn, which is pending LinkedIn’s review, or Telegram. A Monday prompt covering LinkedIn will have no analytics for it. Friday’s prompt explicitly asks the agent to report missing data.

Messages. Our inbox supports DMs only on Instagram. Replies must fall within Meta’s 24-hour window after the customer’s last message, and the agent cannot open a new thread. Comments are readable only on Instagram, Facebook Pages, and Threads.

Rate limits. Platforms cap API publishing, Meta caps Instagram posts per rolling day, and our API has per-minute limits by plan. An agent told to post fifty things will hit one of them. Keep volumes human-sized.

It only knows what you give it. No brand file, no product knowledge, no launch calendar, and the agent fills the gap with generic marketing. The quality of the week above is mostly the quality of the file it reads first.

How much does an AI social media manager cost?

An AI social media manager has two bills: the AI client and the publishing layer. The client cost is your subscription or, if you run the agent from your own code, API usage. PostZen’s plans are on the pricing page; we do not quote prices in posts because they change.

Against a human manager the comparison is qualitative. A person brings taste, relationships, and the ability to notice what the tools cannot see. An agent can speed up repeatable work such as drafting and reporting. Our view is that most teams will want both, with a person responsible for approvals and difficult customer conversations.

The shorter path: one request to every platform

To connect PostZen to your agent, use this connector URL:

https://mcp.postzen.dev/mcp

In claude.ai or Claude Desktop, open Settings, then Connectors, then Add custom connector. Paste the URL and click Connect. In Claude Code, register it with this command, then authenticate through /mcp:

claude mcp add --transport http postzen https://mcp.postzen.dev/mcp

The authorization flow sends you to PostZen to approve access. Start with Read Only and select the profiles the agent needs. Choose Read & Write when you are ready for it to create drafts, remembering that this also permits publishing. Each approval creates a key you can revoke on the API Keys page.

If your agent runs in your own code, this request creates a draft for Instagram, X, and Threads:

curl -X POST https://api.postzen.dev/v1/posts \
  -H "Authorization: Bearer pzn_live_..." \
  -H "Content-Type: application/json" \
  -H "x-request-id: agent-draft-2026-10-20-01" \
  -d '{
    "content": "Draft from the agent. Review before scheduling.",
    "isDraft": true,
    "mediaItems": [{ "url": "https://cdn.example.com/draft-4x5.jpg" }],
    "platforms": [
      { "platform": "instagram", "accountId": "acc_instagram" },
      { "platform": "twitter", "accountId": "acc_x" },
      { "platform": "threads", "accountId": "acc_threads" }
    ]
  }'

A post takes exactly one of publishNow, scheduledFor, isDraft, or queuedFromProfile. The last option puts it in the next free slot of that profile’s queue. The @postzen/node and postzen-sdk Python SDKs wrap the call, the CLI takes postzen auth:set --key ... for scripts, and the MCP docs list every tool with its annotations. Per-platform pages such as the Instagram MCP server explain each network’s capabilities, and our guide to posting everywhere at once shows the same request reaching all 10 platforms.

Connect your existing agent to our social media API and start with Monday’s read-only analytics prompt.