Connect Instagram to Claude with the Windsor.ai connector
Hey, it's Cindy 🌱 This is the live-pipe version: your Instagram numbers flow into Claude on an ongoing basis, so you stop screenshotting your Insights every month.
btw the waitlist for AI System Club is now open! If you're keen on using AI to automate your content, video editing, job search, business or build AI agents, feel free to sign up (only if you want to hehe). Ps. the waitlist gets first dibs:)
One honest thing before you start. Windsor.ai is a third-party connector, not something Anthropic built into Claude. Instagram does not hand its data to Claude directly, so Windsor sits in the middle: it authorises your account, pulls the numbers, and exposes them to Claude. That means a second signup, a second set of permissions, and your Instagram data passing through a company that is not Instagram or Anthropic. It is a legitimate marketing data platform used by a lot of agencies, and it is still a decision worth making on purpose rather than by accident.
the shape of it
What you are actually setting up
Two pieces, and it helps to keep them separate in your head. Windsor is the plumbing, Claude is the thinking. Windsor never analyses anything for you, and Claude never touches Instagram directly.
Windsor.ai, the pipe You authorise Instagram once inside Windsor. It keeps syncing your account and exposes around 50 metrics and 48 dimensions for Instagram (connector page).
Claude, the strategist You add the Windsor connector in Claude, then ask questions in plain English. Claude queries the pipe live, so there is nothing to export, attach or re-upload.
💡 Two different Instagram sources, and you have to pick. Windsor lists Instagram Insights and Instagram Public Data as separate data sources. Insights is your own account (reach, saves, shares, watch time) and needs admin access. Public Data reads any public profile but only externally visible things, so no reach and no saves. For a content plan you want Insights. Note the forever-free plan allows one data source total, so you cannot run both without paying.
the money bit
What it costs, stated plainly
Both sides have a gate, and only one of them is likely to bite you.
🔌 Claude side, effectively free. Web connectors are available across Claude plans, including Free, and Windsor sits in Claude's connector directory rather than being a custom one you build. Free accounts are limited to one custom connector, which is a different thing (Claude help centre).
💳 Windsor side, free with one real catch. There is a 30-day trial (1 user, 10 data sources, 15 accounts), and after it a forever-free plan of 1 user, 1 data source, 1 account. That is exactly enough for one Instagram profile. Paid starts at $23 per month, or $19 per month billed annually (current pricing).
📅 The catch, and it matters for planning. The free plan carries a 30-day history limit: you can only retrieve data from the past 30 days. Paid plans go up to 10 years of history. So on free you are planning from last month, not from your all-time winners, and that changes what you can honestly ask for.
🌱 How I would play it. Use the 30-day trial while it lasts, because that window gives you real history to work with. Run one proper baseline read and save the output as a document. Even after you drop to the free plan and the history shortens, Claude still has that saved baseline to compare against.
from zero
The setup, step by step
📱 Get your account eligible. Instagram Insights only exists for a Business or Creator account, and Windsor authenticates through Facebook, so your account needs to be linked to a Facebook Page you have admin access to. In the Instagram app, open your profile, go to the menu, then settings and privacy, then account type and tools, and switch to a professional account. It is free, reversible, and takes about a minute.
🔑 Create your Windsor account. Go to onboard.windsor.ai and sign up. Use the email you would actually pay with later, because Windsor asks that your subscription email match your onboarding email.
📊 Add Instagram Insights as your data source. Pick Instagram Insights from the source list, authenticate through Facebook when it prompts, and then choose the specific Instagram profile you want connected. If you manage more than one, choose deliberately here, because the free plan gives you one account.
✅ Check the data landed in Windsor first. Before you go anywhere near Claude, confirm Windsor is returning rows for your profile. If the pipe is empty at this end, no prompt in the world will fix it, and you will waste an hour blaming Claude.
🔗 Add the connector inside Claude. Open Customize, then Connectors, then Browse connectors, and search for Windsor. You can also reach the same place mid-chat: click the + button or type /, hover Connectors, then choose Manage connectors. Click connect and sign in to your Windsor account when it prompts.
🔓 Set the permission once. The first time Claude calls the connector it asks for approval. Choosing the always-allow option saves you re-approving on every question. Windsor's own documentation recommends this too, and it is the difference between a smooth chat and twenty permission clicks.
🔎 Sanity-check against something you already know. Ask for your follower count and your five highest-reach posts from the last 14 days, then open Instagram and compare. If the numbers match, the pipe is trustworthy. If they do not, or the profile is wrong, fix that before you build a single plan on top of it.
Read the permission screen properly. Windsor's Claude connector is not read-only across the board: its documentation states the connector can read and write to your connected accounts, and it advertises live actions like pausing campaigns and changing budgets on connected ad platforms. Instagram Insights itself is analytics you are only reading, so the safe habit is to connect the sources you actually need and nothing else.
be precise
What it can and cannot see
This is the part the viral reels skip, and it is the part that decides whether your plan is real. Every field below is in Windsor's published Instagram reference, so you can check me: the field list, or the raw field feed if you want all 100-plus entries.
Per post, the useful stuff Reach, views, saves, shares, likes, comments, total interactions, media type, permalink and the date it went up. This is what a plan should be built from.
Reels specifics Average watch time and total watch time, both in milliseconds, plus reel total interactions. Watch time is the closest thing you get to a retention signal.
What a post did for the account Media follows (accounts that followed after seeing that post) and media profile visits. These two are how you find content that actually recruits rather than content that merely entertains.
Your audience Follower counts by age group, gender, city and country, as lifetime totals. Useful for sense-checking who you are actually talking to.
Five limits worth knowing before you promise yourself a miracle. 1. Media follows and media profile visits are not supported for reels, so your best-performing format is the one where you cannot see follows-per-post. 2. Stories data only exists for 24 hours, so story analysis has to be a habit you run daily or not at all. 3. New followers per day is not available for accounts under 100 followers, and only covers the last 30 days. 4. Follower count has no history, you get today's number, which means follower-growth curves have to be tracked by you over time. 5. There is no field for when your followers are online, so nothing here can tell you the best time to post. Anyone selling you that from this data is guessing.
💡 And the obvious one. None of this shows you the algorithm's internals, anyone else's private numbers, or why a post flopped. It is a disciplined read of your own results, which is genuinely more than most creators ever do, and it is not insider knowledge.
copy these
The 4 prompts that make it a strategist
Run them in order in one conversation. The first one is the setup that makes the other three behave.
🧠 1. the standing brief
You have live access to my Instagram account through the Windsor.ai connector, data source Instagram Insights. From now on, treat that connector as your only source of truth about my performance and never estimate a number you have not actually pulled.
Here is my context: my audience is [who they are and what they want], I post mostly [formats], and what I am building towards is [the goal, for example email signups, a launch, brand deals]. When you analyse anything, always rank posts against my own median rather than against generic benchmarks, and always tell me the sample size you are working from. If a metric I ask for does not exist in the connector's field list, say so plainly instead of substituting a similar one, and if a conclusion is a guess rather than something the data supports, label it as a guess.
📊 2. the baseline read
Pull every post from the last 30 days and build me a table with the permalink, date, media type, reach, views, saves, shares, likes, comments and total interactions for each one.
Then do the maths I would not do by hand. Work out saves per thousand reach and shares per thousand reach for every post, because those two travel further than likes, and rank the posts on that combined score rather than on raw view count. Show me which media types over-index on that ranking and which ones only look good because they got lucky with reach. For reels specifically, add average watch time and tell me whether the high-watch-time reels and the high-save reels are the same posts or two different groups, since that changes what I should make more of.
🌱 3. the 30-day plan
Using only the patterns you found in my actual numbers, build me a 30-day content plan. Every single slot has to trace back to a specific finding, and you should name the post or the pattern it came from in one short line underneath.
For each entry give me the date, the format, the topic, and the opening line I would actually say on camera. Weight the calendar towards the formats and topics that beat my median on the saves and shares score, but include three deliberate experiments in territory I have never tested and label them as experiments so I do not judge them on the same benchmark. Keep the volume realistic for one person doing this alongside a job, and finish with the three numbers I should check at the halfway mark to know whether the plan is working or whether we are wrong.
🔁 4. the monthly re-run
Pull the last 30 days again and compare it against the baseline read we did last month, then tell me what genuinely changed rather than restating the same analysis.
I want three things. First, which of your previous recommendations I actually followed and what happened to those posts specifically, measured against my median at the time. Second, any pattern that has reversed or weakened, because that is the signal I am most likely to miss on my own. Third, one thing I should stop doing entirely this month, chosen because the data supports it and not because it sounds like good advice. Be blunt about anything that is not working.
🌱 Make it permanent. Start a project in Claude, paste prompt 1 into the project instructions, and run everything else inside that project. Now the context is loaded every time and the whole thing becomes a monthly ritual you open rather than a chat you rebuild from scratch.
when it misbehaves
The four things that actually go wrong
Claude says it has no data Almost always the connector was never approved for this chat, or the sync has not run yet. Re-check the connector is toggled on, then ask a dead simple question like your follower count.
The numbers look too short On the free plan you are capped at 30 days of history, so a request for "my best posts ever" quietly becomes "my best posts this month". Ask Claude to state the date range it used, every time.
Impressions come back empty Instagram retired several older metrics and Windsor marks them deprecated, including impressions and video views. Ask for views and reach instead, which are the current fields.
The wrong profile is connected Easy to do if you have a personal and a business account under one login. Fix it in Windsor at the data source, not in Claude, then re-run the sanity check.