Connecting Planable to Claude takes about ten minutes and no code. You add Planable's MCP server as a custom connector in Claude, authorise your Planable account, and describe the work in plain language. Claude drafts posts, files them in the right workspace, and routes them for approval. It creates drafts. It never publishes.
That's the whole build. The rest of this page covers what the setup instructions leave out: what the agent refuses to do, which permissions it inherits, and how you tell whether it's actually working.
What is a social media agent, in plain terms?
A social media agent is a language model that can read a goal, decide which tools to call, act on them, check the result, and keep going until it finishes or needs a human. Anthropic's Building Effective Agents guide, published 19 December 2024, names that loop perceive, reason, act, observe.
A writing tool stops after it produces text. An agent keeps the loop running: it creates the post in Planable, requests approval, watches the status, and takes the next permitted step. One goal in, drafted and submitted posts out.
That difference changes what you're buying. You're not paying for better copy. You're paying for a system that moves work between steps without you in the middle of every handoff.
Here's the honest version. Drafting is the cheap part now. Hootsuite's 2026 statistics roundup reports that 83% of social marketers say AI helps them create significantly more content. More content stopped being the bottleneck the moment that number landed.
What you need before you connect Planable to Claude
Three things: a Planable account, a Claude account, and about ten minutes. The MCP connector is available on every Planable plan, including the free trial, so you don't have to upgrade to test it.
Two connection facts are worth writing down. Your authorisation lasts 180 days, and the live connection expires after 8 hours of inactivity before it refreshes automatically. The second one quietly breaks unattended jobs, and I'll come back to it.
The connector is the easy half. The ready-made skills Planable publishes, which it documents in its connector guide, need a Claude Pro, Max or Teams plan. On the free tier you can still build the same workflow, you'll just be writing the instructions yourself instead of installing them.
How do you build a social media agent with Planable and Claude?
Step one. Open Claude, go to your connector settings, and add https://mcp.planable.io/mcp as a custom connector. On a team or enterprise account, an admin adds it once and it becomes available to everyone.
Step two. Click connect, log in with your Planable account, and authorise. There is no server to run and no API key to paste into a file.
Step three. Ask what you're working with. A first prompt like "list my Planable workspaces and the social pages connected to each one" does two jobs: it teaches you the exact names the agent will use, and it shows you whether your page labels are clear enough to be useful.
Step four. Draft a batch. Give it a brief, a platform, a count and a tone, then tell it to add the posts to the approval queue. Several posts across several channels in a single prompt is the normal case, not a party trick.
Step five. Close the loop. When a reviewer requests changes in Planable, paste the comment back and ask for the revision. The agent updates the draft in place, and the discussion stays inside Planable where the rest of the team can see it.
Step six, and this is optional. Schedule the boring parts. A recurring instruction such as "on the first Monday of each month, pull analytics across my workspaces and send a summary" turns reporting into a standing job instead of a monthly scramble.
How do you keep the agent on brand?
Start a fresh conversation per brand. The context window is per conversation, so a shared session lets one client's voice bleed into another. If you run more than a handful of accounts, that's the most common way this setup goes wrong.
Put the voice guide in your first message: tone, character limits, banned phrases, and two or three sample posts. "Approachable but authoritative, no buzzwords, LinkedIn under 200 words, always end with a question" does far more work than "write in our brand voice".
Name your pages properly. The agent navigates by the labels it can see, so "Client A - LinkedIn" gets targeted precisely while "Social Page 1" produces a guess. Renaming pages in Planable takes five minutes and removes an entire category of errors.
Revise forward instead of restarting. Re-running a task from scratch throws away everything the reviewer liked. Paste the draft back with a narrow instruction, such as rewriting only the first two sentences, and keep the parts that already worked.
The pressure to get this right is not only internal. Sprout Social's 2026 Social Media Content Strategy Report, which surveyed more than 2,300 consumers and 1,200 marketers, found that consumers want human-generated content to be a brand's number one priority, with AI used for audience insight and process efficiency rather than to replace human taste.
Aleyda Solis, writing in Search Engine Journal, makes the operating version of the same point: a personalised editorial workflow is required to keep quality, originality and expertise once AI is producing content at volume.
Where the Planable and Claude setup breaks
It drafts. It does not publish. This is the most important line on the page. Claude creates posts as drafts in Planable, and publishing still runs through Planable's normal approval and scheduling workflow. If you arrived expecting autoposting, you'll be disappointed, and that gap is a feature rather than a missing piece.
Planable's own guide notes that nearly four in five MCP integrations are set up with write permissions. Write access driven by an autonomous loop is exactly why the approval step has to exist.
Permissions are inherited, not granted. The agent can only do what you can do. Without approve rights on a workspace, it can't approve either. Without the Analytics add-on, analytics calls fail. The agent isn't a way around an access limit you already have.
The connection sleeps. After 8 hours of inactivity it expires and refreshes, but a scheduled run fired into a dead session stalls until it reconnects. Keep unattended workflows short enough to recover, and check the first run by hand.
Skills are gated. Four of the six skills Planable ships need the MCP connector, and the other two can run from a CSV export. Using the skills at all needs a paid Claude plan, so the free-tier path is a longer prompt, not a smaller feature set.
Analytics need the add-on. Pulling performance data requires the Analytics add-on on that workspace. Budget for it before you promise anyone a monthly report.
Volume isn't performance. The same Hootsuite roundup reports AI usage among social marketers up almost 180%. That figure tells you what everyone else is doing. It tells you nothing about whether their posts worked.
What should you measure in the first 30 days?
Three numbers, and none of them is the number of posts produced.
Approval latency: how long a draft sits in the review queue. Cycle time: hours from brief to approved post. Revision rounds: how many times a post goes back before it clears.
Track all three by hand for two weeks before you automate the reporting. A drafting agent that cuts cycle time from four days to one earns its seat. One that adds fifty drafts a month to a queue nobody clears has made your week worse, and post volume will happily hide that.
The fourth number, once comments and DMs matter to you, is response time. That's where a human still beats the agent on quality, so measure it to learn where the agent should be allowed to help and where it shouldn't.
When a single-channel agent is not enough
Planable plus Claude solves one job: getting approved social posts out the door. That's a workflow with a narrow scope, and for most people reading this it's the right amount of machinery. Anthropic's own guidance is to use the simplest thing that works and to reach for an agent only when the steps can't be predicted in advance.
It stops being enough when the same brief has to land in four places. A launch message goes to LinkedIn, to an email list, to a landing page and to a paid campaign. Four connectors, four permission sets, and four places where the message drifts. If that's already your week, start with what an AI marketing agent actually handles before you add a fifth connector.
At that point the bottleneck moves from drafting to coordination, and a channel-specific agent can't see the problem because it only ever sees one channel. A full-stack marketing agent is built for that wider loop, and it's a different purchase with a different cost model.
If you're still choosing which social tools deserve a seat, the AI social media tools that actually hold up is the place to start. If the words your reviewers use are the sticking point, the social media marketing glossary is the shorter read.
One last thing, and it's the reason to build this now rather than next quarter. The setup is already trivial, which means the advantage isn't in having an agent. It's in measuring the one thing the agent changes, and in being honest about the parts it still cannot do.