Writing the post is often the easy part. The tedious part comes after: one version for LinkedIn, a shorter one for Threads, a first comment on Instagram, a different time for each, and the video with subtitles. That's the part I wanted an assistant to handle.
This walkthrough connects an MCP client to Postbean, the scheduler I build, so the assistant can create and schedule posts. The same idea applies to any scheduler with an MCP server. Buffer, Postiz, Post Bridge and others have one too, with different tools and permissions.
What the assistant can do
The server exposes the same actions as the web app, as MCP tools:
-
list_accounts,get_workspace: which networks are connected, and the workspace's time zone -
upload_media: images and video, from a URL or, with the local server, from your disk -
create_post_draft,update_post: a default text plus per-network overrides (caption, first comment) -
publish_post,cancel_post,delete_post -
generate_platform_version,generate_campaign: AI rewrites per network, charged as credits -
generate_captioned_video: burns an SRT file into a video
1. Create a scoped API key
In Postbean, go to Settings → Developers and create a key. Pick the scopes deliberately: writing drafts (drafts:write), publishing now (posts:publish) and deleting (posts:delete) are separate permissions. For a first test, leave out the last two. The assistant can then prepare everything, and you press publish yourself.
2. Connect the client
There are two transports with the same tools and the same key.
Hosted (HTTP). Point any client that supports remote MCP servers at:
https://postbean.app/mcp/postbean
Authorization: Bearer <your key>
Local (stdio). Runs on your machine, which matters when you want to upload a file from your disk:
{
"mcpServers": {
"postbean": {
"command": "npx",
"args": ["-y", "@postbean/mcp"],
"env": { "POSTBEAN_API_TOKEN": "<your key>" }
}
}
}
The hosted server reads paths on the server, not on your laptop, so upload_media with a local path only works over stdio. Over HTTP, pass a public URL instead. The setup guide has the exact config for Claude Desktop, Claude Code, Cursor and others.
3. Ask for a post the way you'd ask a person
A prompt I actually use:
Take the release notes in CHANGELOG.md for 1.58. Draft a post for LinkedIn (up to 1,200 characters, no hashtags), a short version for Threads and Bluesky, and add a first comment on Threads with the docs link. Set LinkedIn for tomorrow at 9:00 my time and the others for 13:00, and leave it as a draft.
The assistant calls get_workspace for the time zone, creates one post with three overrides and sets per-network times. The key can't publish, so it stops there. I open the post in Postbean, read it and publish.
4. Two things that tripped me up
Time zones. A model will happily send 17:00Z when you said "5 p.m.". Postbean treats a time without an offset as wall-clock time in the workspace's time zone, and its responses include the offset. Still, check the time on the first few posts.
Length limits. Assistants don't count characters reliably. Each network's limit is checked when the post leaves draft: a 2,400-character caption scheduled to Threads (limit 500) is rejected with a message that names the network and the limit. The fix is a per-network override, not cutting the main text for everyone, and the assistant can apply that fix on its own once it sees the error.
When this is worth it
If you post a few times a week to two networks, a browser tab is fine. It pays off when one piece of content becomes several versions, or when the content already lives where the assistant works: a repo, a changelog, a doc.
Postbean has a free plan with no card. API and MCP access come with paid plans, which start at $5/month; the pricing section has the current details. You can create an account here.
Top comments (0)