Every scheduling tool makes the same assumption: you will sit in a browser, click through a composer, and fill in fields one post at a time. That assumption made sense for a decade. Then agents got good, and the clicking started to feel like friction I had built myself.
I spend most of my working day next to Claude anyway. At some point the obvious question arrived: why am I switching windows to my own product to press buttons an agent could press for me? So Planpost got an MCP server, and scheduling a week of content for my own accounts now looks like a short conversation. This post walks through that workflow as it exists today, including the parts where a human still has to show up.
Why an agent and not a better dashboard
The honest answer is that the dashboard was never the bottleneck. Writing eight platform variants was. Deciding when to post was. Remembering to check what worked was. Those are exactly the chores a language model with tool access is good at, and none of them need a prettier interface. The composer solves the last step, typing a post into a box. The agent collapses the ten steps before it into one paragraph of instructions.
There is also a quieter reason. A dashboard waits for you to show up. An agent can be told "every Monday, look at last week and propose this week," and from then on the job stops depending on my memory.
What MCP is, in one paragraph
MCP, the Model Context Protocol, is a standard that lets an AI agent call tools in outside systems instead of just talking about them. A weather MCP server lets an agent check the forecast. The Planpost MCP server lets an agent operate a social media scheduler: it exposes functions like create_post, generate_carousel, get_best_times, and get_analytics, and the agent calls them on your behalf, authenticated with your API key. Claude supports it, Claude Code supports it, and so does a growing list of other agents. The full tool list lives in the docs.
The setup
Setup is shorter than most OAuth flows. Generate an API key in your Planpost settings, register the planpost-mcp server with your agent of choice, and paste the key into its configuration. In Claude Code that is one command in the terminal. From that moment, "what is scheduled for this week?" becomes a question with a real answer, because the agent can go look.
One habit worth adopting on day one: have the agent create posts as drafts or scheduled items, and keep instant publishing as a deliberate, separate act. The gap between "drafted" and "live" is where mistakes get caught.
A real session
Here is what a normal Monday looks like for me now. I open Claude and write a brief, something like:
"Plan next week for my accounts. The theme is the new analytics page. I want two LinkedIn posts, three posts on X, and one thread that walks through the feature. Link everything to the announcement. Schedule at good times, but nothing on Friday."
Claude checks which accounts are connected, drafts each post in the register of its platform, and asks get_best_times which slots have performed before. Then it shows me the plan: every draft, every platform, every proposed time, in one place. I rewrite two captions because they sound like a press release, approve the rest, and the agent creates the scheduled posts. The whole exchange takes about ten minutes, and most of that is me editing.
Follow-ups work the same way. "Move the Thursday post to next Monday" or "rewrite the third post without the exclamation marks" are single sentences, not a hunt through a calendar UI.
Past plain text
Text posts were the obvious first step. The parts I did not expect to lean on as much:
Carousels. I give Claude an outline and it calls generate_carousel, which turns that outline into designed slides and renders them as images. What used to be an evening in a design tool became a review pass on something that already exists.
Video. There are templates the agent can fill and render for short vertical video. I use this less often, but for announcement clips it beats opening an editor.
Account groups. My accounts are grouped, so "post this to the English accounts" is one instruction instead of five platform pickers.
Bulk weeks. For recurring series there is bulk_create_posts, which takes a whole batch in one call. I use it for formats where the structure repeats and only the substance changes.
The analytics loop
The biggest behavior change came from reviewing, not posting. Each Monday starts with one question: what did last week's posts cause? The agent pulls per-post analytics, and because revenue attribution is wired into my accounts, it can tell me which posts produced clicks and signups rather than just impressions. Revenue attribution is a Pro feature, and I will not pretend otherwise. It is also the reason the Monday brief keeps getting sharper, because next week's plan starts from what last week earned.
The best-time suggestions improve the same way, since they come from when your own audience responded rather than from a generic chart.
Where the human stays
An honest list, because feature posts that read like magic tricks help nobody.
I read every draft before it ships. Agents write plausible copy, and plausible is not the same as sounding like you. Planpost's brand voice profiles narrow that gap, and they are also a Pro feature, but even with one configured the final read is mine.
Approvals still matter. If you run client accounts, a post drafted by an agent can go through the same approval flow as any other: the client gets a link, reviews the post in the browser, and approves without needing an account.
And the agent will not do strategy. Ask it to "grow my audience" and you get generic content at scale, which is worse than no content. The workflow works because the brief carries an actual opinion about what is worth saying. The agent multiplies judgment. It does not supply any.
Finally, some weeks I still open the composer by hand. Seeing a month laid out on a calendar is sometimes faster than describing it in sentences. The agent did not replace the product. It became the fastest door into it.
What it costs and where to start
Planpost has a 7-day free trial, then Creator and Pro plans. The differences are on the pricing page, and the honest summary is that revenue attribution and brand voice sit in Pro. To try the agent workflow: start a trial, connect a couple of accounts, register the MCP server, and ask your agent to plan one ordinary week. The agent page has the shorter marketing version of everything I just described.
The composer is still there whenever I want it. I just notice I want it less every week.