What is an MCP server for social media?

Understand what an MCP server lets an AI assistant do with social accounts, how Postbean implements it, and where permissions still matter.

Published mcp, ai, automation
Illustration: What is an MCP server for social media?

An MCP server for social media gives an AI assistant a defined set of actions it can request from a social publishing service. Those actions might read connected accounts, save a draft, attach media, or schedule an existing post. The server checks the request and returns a result.

That connection changes what a conversation can accomplish. Without a publishing connection, “write next week’s posts” produces text. With access to the appropriate tools, the assistant can also save that text in your workspace. You still need to decide which actions it should take and which credentials it receives.

Postbean implements this connection alongside its REST API. To judge whether it fits your work, start with a single action and follow what happens to it.

The conversation and the stored post

Suppose you ask: “Save a LinkedIn draft about our workshop.” The assistant needs the correct connected account, the workshop facts, and a tool that creates a draft. A fluent paragraph alone supplies none of the account information.

On Postbean’s hosted MCP server, it can call list_accounts, select an account ID, then call create_post_draft with a caption and target account IDs. The resulting draft has a public post ID. Calling get_post reads that stored object back.

If you ask for a correction, update_post edits the existing object. A second call to create_post_draft creates another object. This distinction matters when a conversation contains many revisions: your assistant should retain the ID, not reconstruct your publication history from similar captions.

At each step, ask whether you are looking at a suggestion in chat or a saved result. “I would schedule this for Friday” and “the saved post reports Friday at 09:00” are different statements.

What MCP standardizes

The MCP architecture documentation describes a host application connecting clients to servers that expose capabilities. It separates protocol exchange from how an AI application uses a model. MCP can run locally or through a remote connection.

For your publishing workflow, this means the assistant can discover callable tools and their inputs rather than rely on a paragraph describing an API. It does not mean every server has the same tools, or that a model can bypass the publishing service’s rules.

Treat tool names and schemas as implementation details you must inspect. A tool called “create post” can have very different defaults from a tool called “create draft.” A common protocol does not make those defaults equivalent.

What Postbean exposes

The hosted server includes these practical tool names:

Task Hosted MCP tool
Read workspace time zone get_workspace
Read connected accounts list_accounts
Browse stored posts list_posts
Read one post get_post
Create a draft create_post_draft
Change a saved post update_post
Send an existing post now publish_post
Stop a scheduled post cancel_post

Media has its own actions, including upload_media and get_media. AI tools such as generate_platform_version and generate_campaign are separate from ordinary draft storage and consume Postbean AI credits.

Read the discovered tool schema before constructing inputs. For example, the hosted draft tool uses account_ids; the current local package’s create_post uses social_account_ids. Both refer to selected connected accounts, but copying one payload into the other is not a reliable integration strategy.

Hosted and local mean different file access

Postbean’s hosted MCP endpoint is https://postbean.app/mcp/postbean. It runs on Postbean’s servers and authenticates with a bearer token. That process cannot open a file merely because your assistant knows its path on your computer.

The local alternative is the npm package @postbean/mcp, run through npx with Node 18 or newer. It receives POSTBEAN_API_TOKEN in its environment and talks to Postbean’s REST API. Because the process runs locally, its upload_media can read a local path.

The current local package exposes list_connections and create_post, rather than the hosted account and draft tool names. Always set draft: true for a local draft: without a draft flag or schedule, create_post can request immediate publication.

If your work mostly involves text and reachable media URLs, the hosted connection avoids running a package on your machine. If your assistant handles local exports, the local package can take care of transferring those files. Choose based on the workflow you actually have.

Permissions exist outside the prompt

Postbean keys carry selected scopes. Reading connected accounts uses accounts:read; reading posts uses posts:read; writing drafts uses drafts:write. Immediate publication and deletion have separate scopes, posts:publish and posts:delete.

The account behind the key must also be allowed to perform the action. For post edits, authors and workspace managers have different relationships to a post than another ordinary member. A tool connection does not turn every workspace member into an administrator.

Prompts remain useful for expressing intent: “save drafts only” is clear. They are not a substitute for limited credentials or a reviewed publishing process. Scheduling is itself a future send, so account for schedule-changing actions when designing your boundaries.

When a permission error appears, changing the wording of the request is not a fix. Check the key and user access. Moving from MCP to REST uses the same permission system rather than providing an alternate route around it.

A tool result is not always a completed publication

Some work continues after an initial response. On hosted MCP, uploading media from a URL reserves the media and downloads it in the background. A response with ingest_status: pending means you should inspect that media, not upload the same URL repeatedly.

Scheduled posts wait for their selected time. Publication can also produce different outcomes across target accounts. A response that successfully saved a post does not establish that every network accepted it later.

Build your operating instructions around these stages. Ask for the saved ID after creation, the saved timestamp after scheduling, and the account-level result after sending. Keep error messages with the corresponding object so someone can investigate without guessing which version was attempted.

Time and platform rules still apply

An assistant needs an actual clock context. Postbean’s hosted get_workspace returns the workspace’s time zone and local time. Offset-free timestamps are interpreted in that zone; timestamps carrying an offset or Z identify a particular instant.

This is why “tomorrow morning” should become an explicit date and time before you approve a batch. A model should not append Z to a local hour just to make the string look technical.

Platform-specific requirements also remain. Postbean reports text-limit problems in publish_issues, and scheduling or publication checks those limits. For TikTok, get_tiktok_creator_info supplies account-specific audience options. The protocol does not remove these decisions or make guessed values acceptable.

Decide whether you need MCP or REST

Use MCP when a person is directing an assistant through changing instructions: choose an account, revise a caption, attach this asset, and inspect a result. The assistant can select tools while you work through the task.

Use REST when your application already knows exactly what to do. A service that receives an approved article and creates a corresponding draft can call a documented endpoint directly. Adding a language model to that fixed operation may add another decision point without helping.

You can use both approaches in the same business. An automation can prepare drafts while an assistant helps an editor inspect them. Keep the same post IDs across both so revisions remain attached to the right object.

Before expanding access, run one complete test with a draft and read it back. Then try a deliberate validation failure, such as an overlong network version, and confirm that your process surfaces the error. The connection guide covers installation, while the developer API guide covers direct requests.

Sources

Write it once. Publish it everywhere.

Connect your networks and schedule your first post in a few minutes. Free plan, no card.

Create your free account