Claude Cowork for Social Media: Research, Drafts, Audits and a Weekly Routine

Claude Cowork is Anthropic's agent for work that is not coding. Anthropic's getting started article says it brings Claude Code's agentic abilities to knowledge work with no terminal involved: instead of answering one prompt at a time, Claude carries out a task with many steps. The Cowork overview in Claude's docs adds that you describe an outcome and come back later to finished work, such as documents, organized files or a piece of research, and that Claude can split a big job into parallel workstreams.

The part that matters for social media is where it works. On the desktop app, Claude reads and writes files on your own computer without uploads, but only inside folders you connect. It can use connectors to other services, a built in browser that opens sites, reads pages, clicks and types, and plugins that bundle skills, connectors and helpers into one install. The getting started article says sessions run in the cloud in an isolated environment and keep going when your laptop is closed, while anything that touches local files, the browser or other apps needs the desktop app open and connected.

The name is in the middle of a change. On September 16, 2026, Anthropic announced that Cowork and chat are merging into one Claude, so you no longer choose between a chat and a task: Claude decides what a request needs. Pro and Max get the merged app first, over a few weeks, with Team and Free later. Your chats, projects, connectors and skills carry over. Anthropic's help articles and the Settings menu still say Cowork, and people still search for it by that name, so this page does too.

This page is about using that agent for social media. If you mostly chat with Claude and want the wider picture of Projects, Skills and prompts, read Claude for social media marketing. If you ship software and want posts to start from your repository, read Claude Code for social media. Here we stay with what only Cowork brings: folders on your computer, connectors with per tool permissions, and scheduled tasks that run without you.

"How to use claude cowork for social media" is the most common search on this topic, and the honest answer is that Cowork is a strong assistant for the slow parts of social media and a weak publisher. The slow parts are reading, sorting and writing: going through a month of release notes to find the three changes people will care about, comparing last quarter's posts to see what worked, writing ten variants of an ad, turning one launch into a week of posts for different networks. Cowork does these well because it can open the actual files rather than wait for you to paste them in.

The weak part is getting a finished post onto a network at the right time. Cowork has no social media accounts of its own and no posting feature. What it has is connectors, and a connector to a social media tool is what turns a folder of drafts into posts on a calendar. Anthropic's Cowork product page says Claude uses connectors first and falls back to the browser or the screen only when it needs to, so with the right connector it calls the tool instead of clicking around a website.

So a working setup has three layers. A folder (or a Cowork project) that holds your raw material: release notes, App Store text, screenshots, past posts, an exported analytics file and a short voice guide. Cowork itself, which reads that folder and writes. And a social media connector, which designs the picture, writes a caption for each network, saves the post as a draft and, if you allow it, schedules it. The rest of this page goes through each layer, then puts them together in a weekly scheduled task.

Yes for the words, partly for the pictures. Cowork writes captions, hooks, hashtags, thread outlines and carousel scripts very well, and it saves them as files you can keep. The getting started article lists its outputs as Excel files with working formulas, slide decks and documents, with Markdown drafts you can edit in place. A social media manager can work entirely in those formats: a spreadsheet of posts, a document of captions, a deck that shows a carousel slide by slide.

The finished picture is the harder part. A post on Instagram is an image of a fixed size in your brand's colors and fonts, often with a real screenshot of your app inside a phone frame. Cowork can describe that slide perfectly and can mock up a layout, but turning it into a set of matching, correctly sized JPEG files every week is design work it is not built for. That is the gap a design connector fills.

Spreadcast is one such connector, made for people who ship App Store apps. Connected to Cowork, it gives Claude tools to design posts in 11 kinds and carousels of 3 to 8 slides at 1080x1350 pixels, in your app's brand kit, with your real App Store screenshots in an iPhone frame. It makes no AI photos and no video. Claude writes the words; Spreadcast draws the post and saves it in your account as a draft you approve.

Not by itself, and that is by design. Cowork does not log in to Instagram or LinkedIn for you, and it has no publish button. Searches for "can claude cowork post on social media" usually come from people who saw it handle email and files and expected the same for posts.

There is one way it technically could: the built in browser can open a website where you are signed in and click buttons. Anthropic's article on using Cowork safely is cautious about exactly this kind of use. It asks you to be deliberate about which sites Claude works in, especially ones where you are signed in, to switch to manual approval for accounts and for actions that are hard to undo, and to avoid scheduled tasks that send messages on your behalf. It also lists content you publish and messages sent for you among the things you remain responsible for. A post clicked out through a browser is all of those at once.

The clean route is a connector. A social media tool connects to each network through that network's official sign in, holds the tokens on its side, and gives Claude a short list of tools such as design a post, save a draft and read results. Cowork calls those tools; the tool does the publishing, at the time you approved, with the network's own rules applied. Claude never sees your Instagram password, and every post leaves a record in the tool.

With Spreadcast, a connection made with Make Drafts can design posts, write captions, check them and save drafts, but it cannot schedule or publish. A connection made with Make and Publish, on Spreadcast Pro or Studio, can schedule or publish once you say yes in the chat. For anything that runs unattended, keep the first kind.

Through a connector, yes; directly, no. "Claude cowork instagram connector" is a real search, and there is no Instagram connector from Anthropic. What exists are social media tools whose connectors post to Instagram for you, inside the limits Meta sets for every outside app.

Those limits come from Meta's content publishing guide. It covers publishing to Instagram professional accounts, so a personal account has to switch first. It limits each account to 100 posts published through the API in a moving 24 hour window, and counts a carousel as one post. JPEG is the only image format it accepts, a carousel holds up to 10 images or videos, and Meta fetches each file from a public address at the moment of publishing. That last rule is why a picture on your laptop cannot go straight to Instagram: a tool has to host it first.

Spreadcast handles those parts. It saves its designs as JPEG pictures, hosts them, and checks each post against Instagram's rules before saving. When Cowork has a new screenshot in your folder that is not in the brand kit yet, Spreadcast's upload tool takes a small picture sent directly or a public link. Results come back as likes, comments and follower counts per network; Instagram reach is not among them.

The same pattern covers the other networks people search for, such as "claude cowork linkedin connector" or a Telegram connector. Spreadcast posts to Instagram, Facebook Pages, LinkedIn (your personal profile), Threads, Pinterest, TikTok, YouTube, Bluesky, Telegram channels, Discord channels and Mastodon. TikTok posts stay private until TikTok finishes auditing the app. For the details of each, see the Instagram MCP server and LinkedIn MCP server pages.

A connector is an MCP server, a service that gives Claude a set of tools for one product. The Cowork overview says you manage connectors, skills and plugins from Customize in the sidebar, and that Cowork loads the ones enabled on your claude.ai account at the start of each session. It does not read the folder where the Claude Code command line keeps its own settings, so a server you added in the terminal does not appear in Cowork until you add it in Customize too.

There are two kinds. Connectors in Anthropic's directory install with a click. Anything else, including most social media tools, you add as a custom connector by its address. Claude's guide to adding a connector that is not in the directory says a remote server can be added by URL on Free, Pro, Max, Team and Enterprise plans, with one custom connector on Free. On Team and Enterprise an Owner adds it for the organization and each member then connects with their own account. Sign in usually goes through the tool's own OAuth screen; a fixed key sent as a request header is in beta for a limited set of organizations.

Once connected, each tool gets its own setting. The getting started article lists three: Always allow, Needs approval and Blocked. You can also switch a whole connector off for one conversation from the plus menu. On Team and Enterprise, Anthropic's admin article says members cannot skip approval for tools that write unless an Owner turns that on, and that a tool counts as read only only when the connector marks it so. It adds that most custom connectors do not mark their tools, so every tool on them asks each time. Spreadcast marks its reading tools as read only, so lookups such as your brand kit or last week's results do not stop the task.

The same guide is plain about the risk. A remote server can read data, create, change or delete it, and act on your behalf, so connect only servers from organizations you trust, read the permissions on the sign in screen, and block the tools you do not need.

Does Claude Cowork have scheduled tasks? Yes, and they are what turn it from a helper you visit into a routine. Anthropic's scheduled tasks article says they are on every paid plan. Claude saves your prompt as the task's instructions and runs it on the schedule you pick: hourly, daily, weekly, on weekdays, or by hand. In the merged app you can simply describe the timing, such as every Monday at 9 AM. Each run is its own session that you review when it is ready, and the Scheduled page in the sidebar lists upcoming and past runs.

You can start one in three ways: type /schedule in a conversation, let Claude build it from Scheduled, New task, Create with Claude after a few questions, or set it up by hand with a name, the prompt, an approval mode, a frequency and, if you like, a model and a working folder. The article says a scheduled task has the same abilities as a normal Cowork task, including connected tools, skills and installed plugins.

Where it runs decides whether it runs. Tasks run remotely, so they keep their schedule while your computer is asleep or the desktop app is closed, using your connectors and files saved to your Claude account. Tasks that need local files or apps run only on your computer. Pro and Max tasks created before October 6 that ran on your computer keep running there.

People also search for a scheduled tasks limit and for what happens when a task was due while the computer slept. Anthropic's article lists neither a cap on the number of tasks nor rules for missed runs, so treat the practical limit as your plan's usage: the getting started article says every step of a task uses tokens and that tasks with many steps use more than a quick question. For a weekly social media task, the clean answer to both questions is to keep it in the cloud: put the inputs in your Claude account or behind connectors, not in a local folder.

A scheduled task runs while nobody watches, so it should only be able to do things you would not mind finding on Monday. Anthropic's safety article makes the same point in general terms: avoid scheduled tasks that send messages on your behalf or take actions that are hard to undo, and stay close to tasks with real world consequences. A public post is both.

The approval mode you pick when you set up a task by hand matters here. A mode that asks before every write may stop the run and wait for you, while a mode that skips approvals lets the task use every tool you left open. A drafts only connection removes the dilemma: let it save drafts freely, because a draft goes nowhere until you approve it. Spreadcast's own instructions also tell Claude to show you the post, the networks and the time before any schedule or publish, even on a connection that may publish.

There is a quieter reason too. Social media rewards consistency more than speed, and a batch of drafts reviewed in ten minutes on Monday beats a stream of posts that each need checking after they are already live. You keep the judgment, Claude keeps the routine.

"Claude cowork social media audit" is a common search because an audit is the job Cowork fits best. If your network's dashboard lets you download post data as a spreadsheet, Cowork can read it the way an analyst does. The Cowork product page lists CSV and Excel among the file types it works with, and its own example is a weekly marketing readout pulled from analytics, compared week over week.

Start with the right question. "Audit my Instagram" gets a generic answer. "Using the file in this folder, tell me which post formats got the most comments per post over the last 90 days, which weekdays and hours did best, which topics fell flat, and three things to change next month, with the rows behind each point" gets a report you can act on. Ask for the output as a spreadsheet with a summary tab, so you can check the sums yourself.

Watch the numbers. Claude should compute from the file, not estimate, and every figure in the report should trace back to a row. If the export is missing a column, such as reach or saves, the right answer is "the file does not include this", not a guess. Put that rule in the request.

An audit from a file is a snapshot. If you want the same questions answered every week without exporting anything, a connector that reads results directly is simpler: Spreadcast's analytics tool returns totals per network, top posts, best hours and weekdays, follower counts and your App Store ratings, which a scheduled task can read before it plans the next week.

Searches for "claude cowork social media ad" are mostly about copy, and Cowork is good at it. Give it the product, the audience, two or three angles you want to test and the placement, and ask for a spreadsheet with one row per variant: primary text, headline, the angle it tests and the image it needs. It can also read a CSV of past ad results and tell you which angle won.

Give it the placement's limits up front. Meta's ads guide for Instagram feed image ads recommends primary text of 125 characters and a headline of 40, allows up to 30 hashtags, accepts JPG or PNG up to 30 MB, and recommends a 4:5 ratio at 1440 by 1800 pixels. Ask Claude to count characters for every row and flag any that run long.

What it should not do is run the ads. Cowork can draft, but launching a campaign means spending money from an ad account, and Anthropic's safety article lists purchases and financial transactions among the things you remain responsible for. Keep ad accounts out of connectors you give a scheduled task, and launch campaigns yourself in the ad tool.

For the creative itself, the same design connector helps. Spreadcast designs pictures in your brand at 1080x1350, a 4:5 ratio that fits that placement, with your real screenshots in an iPhone frame; it does not make video ads or AI photos. If ad design is your main need, our ad creative generator page covers that side in more depth.

"Claude cowork social media skills" and "claude cowork social media manager" are searches for the same thing: a way to make Cowork behave like a social media manager every time, not only when you write a long prompt. Cowork has three tools for that.

Skills teach Claude one repeatable job, such as turning release notes into a carousel outline or writing a founder post for LinkedIn. Plugins bundle skills with connectors, helpers and hooks into one install. Claude's plugin guide for Cowork says you install them under Customize, Plugins, from Anthropic's catalog, your organization or a file, and that a plugin you install is saved to your account, so its skills and connectors also work in chat and in Claude Code. Installing a plugin does not sign you in to its connectors; you connect each one afterwards. Anthropic's Cowork page lists a Brand voice plugin among its examples.

Projects hold the context. Claude's projects guide describes a Cowork project as a bundle of local folders, standing instructions, links and a memory of its own that persists across sessions. One project per app works well for social media: the folder with notes and screenshots, instructions with your voice rules, and a memory that learns which posts you approved. Two details matter. Cowork projects live on your computer only and are not shared, unlike claude.ai projects, which can be shared on Team and Enterprise. And Claude reads single files up to 50 MB.

If you want to hand off a bigger job and check back later, Dispatch runs it in the background on Pro or Max with the desktop app. It splits the work into tasks and can route them into a project. Permission requests are forwarded to you, and one left unanswered for ten minutes is denied while the task carries on without that action, which is another reason to keep the social media connector on drafts.

If your raw material is documents, spreadsheets and folders, and you want a routine that runs every week, use Cowork. That covers most founders and marketers who are not engineers. If your posts start from code (release tags, a changelog, pull requests), Claude Code for social media gives you hooks, permission rules and routines triggered by a release, which Cowork does not. If you mostly want a good writing partner and a place to keep your voice, the chat app with a project is enough, and Claude for social media marketing walks through it.

They are not exclusive. A connector you add on claude.ai shows up in Cowork and in the chat app, and a plugin installed in Cowork also brings its skills to chat and Claude Code. Many people draft in chat, run the weekly batch as a Cowork task and keep Claude Code for launch days.

Less than you might fear, more than a chat. A weekly drafts task reads a few files or tool results, calls a handful of tools and writes a few hundred words, which is light next to a task that rebuilds a spreadsheet of thousands of rows. Anthropic does not publish a number per task, and nobody can give you an exact one for your setup, so watch it for two weeks under Settings, Usage.

A few habits keep it down. Keep the instructions short and specific. Point the task at the files it needs, not your whole drive. Leave connectors the task does not use switched off. And know that Auto mode, which runs a safety check on every action, uses more of your limit than the other modes. For most solo founders, Claude Pro covers a weekly task plus daily chat; Max makes sense if Cowork also does your research and reporting all week.

For the connection itself, the Spreadcast MCP page and the social media MCP server reference list every tool Claude gets and how sign in works. If you are comparing connectors, the best social media MCP servers ranks the options with prices, and the guide on how to schedule social media posts with Claude is the shortest path to a first scheduled post.

And if the hard part is deciding what to post at all, how to promote an app on social media covers the content a weekly Cowork task will then help you produce.