How-To Guide
How do you use Claude Cowork for social media?
Claude Cowork folded into the Claude app: which surfaces run the weekly social loop, what a connector needs, and where the human editor stays.
By the Superdirector editorial team. Four official Anthropic pages, two practitioner pages and three competitor guides were read November 14, 2026. Where a source is a vendor's own page, we say so.
At a glance
How do you use Claude Cowork for social media?
You use Claude Cowork for social media by naming a goal, granting it a folder to read and the connectors it may call, and reviewing what it drafts, while a scheduled task keeps the loop running when you are away. Cowork is the agentic side of Claude: it reads and writes files, opens a browser for web work, and can run a task on any cadence unattended. It cannot touch your social platforms on its own. Claude holds no social platform credentials of its own, so every publish, comment or message goes through a connector you connect and set permissions on. The practical split is this: let the agent write, plan and organize, and let the connector hold the queue, then approve the settings once instead of every post. If you would rather hand the whole account to a service, the Superdirector Growth plan at $199 a month covers scheduled research, routines and supported Instagram and X publishing once posting is turned on for an account, and Starter at $29 a month is the hands-on plan for people who drive the work from Claude or ChatGPT themselves.
- Cowork reads the folders you grant and finishes multi-step tasks, and Anthropic documents that a task on a cadence runs unattended.
- Publishing always runs through a connector. The Verified, Community or Custom label on a connector says how closely Anthropic reviewed it, not whether it will post well.
- Approval is a setting you choose before the first job: Manual asks before each publish or schedule call, Auto keeps going. Anthropic's small-business package states the boundary plainly: you approve before anything sends, posts, or pays.
- Agentic work consumes Claude's rate limits faster than chat, so plan the weekly loop around the limits, not around one big batch.
A fork narrows the surface question to files and cadence, or a code project. Original diagram from the official Cowork-versus-Claude-Code FAQ and the help-center note that connectors are shared across surfaces, read November 14, 2026.
Make it work for your brand
Find it. Remake it. Post it. Superdirector turns the viral formats in your niche into your brand's carousels, X posts and video scripts, and keeps the calendar full.
The short answer: an agent for the work between drafts and publishing
Most social media guides that mention Claude stop at the draft. The draft is the easy part. What makes an account run is everything around it: pulling what performed last week, choosing what to publish next, keeping research organized, and moving finished work onto a calendar without thirty clicks. Claude Cowork is the part of Claude built for exactly that middle layer. You give it a goal, it works across the files and tools you chose, and the official product page describes the shape briefly: "Give it a goal, and it works across your files and tools."
The boundary to understand before anything else is that Cowork is not a social tool. Anthropic's custom-connector guide describes the handoff: you authorize the app through OAuth so Claude works on your behalf "without Claude ever seeing your actual password", and every publish, comment and direct message happens through a connector, a tool you connect once and set permissions on. What Cowork itself brings is the reading, planning and organizing, plus two capabilities that change the weekly rhythm: Anthropic documents that a task can be put on any cadence and run unattended, and it opens its own browser for web work so research does not hijack your laptop.
This guide walks the setup in the order an operator meets it: which Claude surface to use, how to connect a publishing layer, what to ask first, and where publishing gets its approval. Every external claim links to the page we read it on, and every number that is a vendor's own claim is labeled as theirs, next to the pages we read.
What Cowork is, and what changed in 2026
Cowork launched January 12, 2026 as a research preview that Anthropic described as Claude Code for the rest of your work, according to Simon Willison's first-impressions post that day, and it was inside the Claude desktop app for Max subscribers at launch, with Pro accounts added within the week. Anthropic describes Cowork's difference from Claude Code in its own FAQ: Claude Code is built for software engineering, while "Cowork is built for non-coding knowledge work". One practitioner reading worth knowing: what ships as Cowork is essentially Claude Code with a friendlier interface and a filesystem sandbox already set up for you.
Since then, the product folded into the main Claude app. Anthropic's product page now leads with "Claude Cowork is now just Claude. Rolling out to Pro, Max, and Ultra, with more plans to follow", and the unified experience rolls out plan by plan, meaning some accounts still see separate Chat and Cowork options. Practitioner pages observed the fold-in reaching paid plans from September 2026 onward, though no official page we read documents the exact rollout order, so the safe check is your own account, not the plan name.
Two mechanical facts shape the social media workflow. First, agentic tasks cost more of your rate limits than plain chat: Anthropic says "Claude Cowork consumes limits faster than Chat", so one heavy audit task can eat the same quota as a week of writing. Second, scheduled work has two modes worth telling apart. A cloud task runs even when your laptop is closed, and Anthropic documents that Cadence tasks keep running unattended; a task that reads a folder on your own machine needs the desktop app open, which is why folders belong in scheduled desk sessions and research belongs in cloud tasks.
What Cowork plugins are, and which ones a social team uses
Anthropic now presents plugins as the customization layer for Claude and for Cowork. Its product page says plugins "bundle your tools, knowledge, and workflows into a single install so Claude shows up as a specialist from day one", and that Cowork can be customized with plugins that "bundle any skills, connectors, and sub-agents together". That definition matters for a social team because a plugin is not a second kind of connector. It is a package that can carry the connector you already approved, plus the skills and instructions that decide how the agent uses it.
The menu names are moving at the same time. Blotato's setup guide tells readers the connector sits "under Customize (being renamed Plugins), then Connectors", so a guide that says Customize and a guide that says Plugins can describe the same screen (Blotato's Claude setups). If your account already shows Plugins, follow that name; if it still shows Customize, the step order is unchanged.
Which plugins a social team actually uses comes down to two jobs. A publishing plugin carries the connector that puts a post live. Blotato's own guide walks the user through "Plugins > Add Plugin", a skills repository and "Blotato Content Creation Skills", and Upload-Post says its connector "works with connectors and plugins, the same way Claude chat does", with a separate plugin for Claude Code (Upload-Post's Cowork guide). A voice plugin shapes the writing: Anthropic lists a Brand voice plugin that analyzes "existing documents, marketing materials, and conversations" to distill "your brand's voice" into enforceable guidelines. For a team, the product page also records an April 9, 2026 addition, a private plugin marketplace that admins can create.
The rule for a brand account is install only what the declared capability supports. Read a plugin's tool list the way you read a connector's, keep the post and send tools behind Needs approval until the plugin has earned them, and prefer the private marketplace when a team needs a reviewed set instead of an open directory.
Decide which Claude surface fits your account
Three Claude surfaces can hold the same social media work, and they differ in where the files and the publishing live. Plain chat on the web works for writing posts when you paste the content in yourself. Cowork reads the folders you grant on the desktop app, opens its own browser for web research, and runs scheduled tasks. Claude Code runs in your terminal and suits jobs tied to a code project, like publishing release notes from a repository, and the heavier pipeline work that one practitioner kept in Code after Cowork refused the tooling.
The connector belongs to your account, not to one surface. Anthropic's help center states "Custom connectors using remote MCP are available on Claude, Cowork, and Claude Desktop for users on Free, Pro, Max, Ultra, Team, and Enterprise plans", so the connection you add once is available where the work starts. The choice that matters for a brand account is between driving the work yourself and having the queue run on a schedule. A solo operator who reviews everything by hand can live in a chat. A brand running accounts on real calendars needs the folder access, the recurring task and the queue that keeps publishing when the conversation has long ended.
Set up Cowork the right way before the first job
Setup has four moves, and doing them in order is what keeps the first task clean.
- Pick the accounts' home. Connect your social accounts in a scheduling or publishing tool that holds them with OAuth sign-in, so Claude never touches a platform password. This is the pattern both practitioner connectors document: the connector layer holds the platform connections, and OAuth means no key to paste.
- Add the connector in Claude. Open Customize (being renamed Plugins), then Connectors, then the directory, and connect the tool from its listing; or choose Add custom connector there and paste the MCP URL the tool documents. Anthropic's help center notes that free accounts can add exactly one custom connector this way.
- Check the tier before you rely on it. The directory shows how closely Anthropic reviewed a connector: Verified, Community or Custom. Anthropic's verification page says verification is not a security audit or a guarantee of how the connector will perform, and that a connected Community connector's tools carry the same capabilities and access as any connector you grant.
- Verify with a read-only prompt. Ask Claude to list the accounts the connector can reach. If the list comes back complete, the connection is real; if a call fails with a subscription or permission error, the connector's own plan side is usually the cause, not Claude.
One more setting matters at this point. On Claude Team and Enterprise, an owner manages whether members can connect tools in the first place, so if the directory is missing or blocked on a work account, ask the owner before you redesign your workflow around a tool you cannot reach.
Give the first task and read what the agent plans
The first job should produce something you would have written down anyway, so you can grade the agent against real work. A weekly review is the job the official pages use themselves: Anthropic's own example, from its small-business launch, runs "analyze your HubSpot campaign performance, draft the promo strategy, and generate the assets in Canva to prepare your next send", a research task and a draft task, with publishing left out of the first job. On the same pattern, a founder's first Cowork task looks like this: read the folder where this week's analytics screenshots sit, list what performed above your account's usual level, and propose three angles for next week in your own voice.
Read the plan before the work. Anthropic documents a plan-first approval mode where Claude shows its plan and waits for your approval before anything significant, and the product page adds that the agent keeps its screen actions as a last resort: "It opens browsers for web tasks and takes over your screen only when it has to." For a first job, ask the agent to show the plan and stop, then approve, so you learn what it does with your folders without any side effects. Note Anthropic's warning that agentic tasks consume rate limits faster than chat: budget the review tasks like a scarce resource, not a background noise.
Feed what you learn back immediately. Both long-running practitioner setups converge on the same habit: the brief, the voice guide and the standing instructions are living documents, and the sessions that end with the brief updated are the ones that keep the writing on-brand. This is the ordinary maintenance cost of an agent you steer, and skipping it is why the drafts drift after week three.
Turn on publishing with an approval you chose
Publishing through a connector is two decisions at once: who can call the post tools, and what those tools do when called. Claude's own permission model, as the connector roundup documents, offers "Always allow, Needs approval, or Blocked" for each tool under a connector's permissions, and practitioner guides set the publish and schedule tools to ask each time until the pattern is trusted. Anthropic's verification page adds the tier context: a Verified connector has been tested for quality and compatibility, while Anthropic "screens Community connectors before listing them" without an in-depth review, so the badge says who did the review and you say who is accountable for the posting.
Practitioner setups put the same request in the task prompt, which is a habit worth copying. Upload-Post's Cowork guide instructs the agent to show the caption, the accounts and the date and time, and wait for an OK before it publishes, schedules, replies or sends a DM, and to check the upload status after publishing so you get the result for each platform. This prompt-level approval stacks with, not replaces, the tool permission, and it survives the moment the agent overreaches on a confusing request.
One Anthropic warning belongs in this section because it changes what Research means: a connector roundup quoting the custom-connector guide notes that during Research, "Claude can invoke tools from your connectors automatically without further approval". If you run Research with a queue-full of connectors live, block the post and send tools first. The queue itself should live in the connector: scheduling runs in the connected tool, so a queued week keeps publishing whether or not the chat or your laptop stays open, and the agent can read the queue back to you, move a slot, or batch-edit captions next Monday.
| Connector tier | What Anthropic says about it | What it means for your posting |
|---|---|---|
| Verified | Tested tools for quality and compatibility; not a security audit or a guarantee of performance | The stricter tier, still test one post before loading a week |
| Community | Screened before listing, not reviewed in depth; reminder shown before connecting | Widely used by publishing tools; set publishing tools to ask at first |
| Custom | You added it yourself; Anthropic has not reviewed it | You carry the review; keep it out of unattended Research |
Keep a weekly loop and grow it deliberately
A loop survives its first week when it is scheduled, reviewed and measured. Put the review on a cadence: Anthropic's product page documents that a task can be scheduled for any cadence and runs unattended, which is how the week's reads are waiting on Monday morning without you opening anything. Drafts then move through the review you set up: read the captions, check the accounts and the times, and let the connector queue them. The example below is a pattern, not a real account, and no observed performance data is behind it.
Measure the loop the way you would any process: time to produce the weekly set, share of drafts you shipped without edits, and how often the agent misread a source so badly that a human had to fix the plan. Grow the account by hours saved and shipped posts, not by the number of connectors attached. When the volume outgrows you, that is not a prompt problem; it is usually the point at which hands-on operation starts to compete with delegation, which is the decision in the last section.
A five-station loop puts the scheduled read, your review and the connector queue in the order the section describes; it is an illustrative pattern, not a real account. Original diagram from Anthropic's product page and the loop this section sets out.
Where the hands-on job ends and an operated account starts
Running Cowork for social media is a real job, and it is yours: grant folders, connect tools, approve plans, read drafts weekly, and keep the brief alive. That is the work of a half-day a week for one account, more for three. The honest test is not whether you can run the loop, it is whether you want it on your calendar. Two ways to buy the same outcome separate on that line.
If your job is steering, Superdirector's Starter plan at $29 a month is the hands-on option: the pricing page puts it plainly: "If you work from Claude or ChatGPT, Starter at $29/month is all you need: live social data, ideas, scripts and downloads." Starter connects your AI to Superdirector's MCP, so the tool calls come from wherever you work. If the account should run without you, the Growth plan at $199 a month is the operated job: scheduled research, routines and, on Growth, Superdirector "publishes eligible Instagram carousels and single X text posts by themselves once posting is turned on for an account". The buyer difference is not features, it is where the approval habit sits: on the operated plan, posting starts with one warning that names the account, the public effect and the cost, and the customer keeps the decisions on what posts, what changes and what gets paid.
Both paths keep you as the publisher, and the pricing page says so directly: "You stay the publisher. We sell business growth, not guaranteed results." AI video and fully autonomous channels are not available yet, so any page that promises either today is ahead of the product. If you are comparing against doing it all from the desktop app, the deciding question is the weekly review: an agent with a folder and a connector still needs a reader, and whether that reader is you is what the two prices are pricing.
Two panels place where approval sits at the centre of the buying choice. Original diagram from Superdirector's pricing page, which also marks AI video and full autonomy as not yet available, read November 14, 2026.