Build Logs · 19 Oct 2026 · 6 min
Plug your AI into the apps you already use: ten connectors worth setting up, by job
Gmail and Calendar are the obvious ones. Ten less obvious connectors by job, and the real ways they went wrong for me.

A friend asked me last week why her AI “never knows anything”. She was pasting emails into it one at a time, then pasting the reply back into Gmail. Mine reads the inbox, drafts the replies and leaves them in Drafts for me to send. Same tool. The difference is one permission she hadn’t switched on.
That permission is called a connector. It lets your AI read and use an app you already have, under your own login. In Claude it lives under Settings, then Connectors: pick the app, sign in once, and any chat can use it. The AI never sees your password, and it can only see what your login can see.
You already know about Gmail and Calendar. Connect those first if you haven’t. This list is the less obvious stuff, the connectors I wish someone had pointed me to earlier, organised by the job you do.
The ten, by job
1. Google Drive and Sheets. Anyone who runs a process. A shared folder is the AI’s memory, and a spreadsheet is its control panel. My content pipeline is a Drive folder with a spec file, a voice guide and four stage folders, plus a sheet with one row per piece, a status column and a comments column where I type what’s wrong. The AI reads it every morning and acts on it. If a decision isn’t in the folder, it didn’t happen.
2. Windsor.ai. Marketers and anyone who pays for ads. It pulls numbers from ad platforms, analytics and a few hundred other sources into one place. Instead of opening five dashboards, you ask “which campaign spent the most last week and got the least back?” and get one answer. It can also change things on some ad platforms, so keep it on read-only until you trust it.
3. Metricool. Social media managers and small business owners. Schedules posts, shows which networks are performing and suggests the best time to post. It’s how the store’s videos get scheduled. The AI writes the caption, attaches the clip and picks the slot. I check it before it goes anywhere.
4. TipRanks. Anyone who follows stocks. Analyst ratings, price targets, earnings dates and company financials. Handy for “what’s reporting this week from the companies I own, and what did analysts say last time?” It’s data, not advice. Treat anything it tells you as a starting point for your own thinking.
5. Shopify. Anyone with an online store. Products, prices, stock levels, orders, discounts. It runs my store day to day: about 40 products live, shipping to 18 countries, with me approving spend and anything customer-facing. My two earlier stores were built before AI and each took weeks. This one took a week, with the connector doing most of the clicking.
6. Higgsfield. Anyone who needs images or short video. Generates pictures and clips from a description. Good for product shots, social posts and blog heroes. Read the credit cost before you start: video is far more expensive than images.
7. vidIQ. YouTube creators. Channel analytics, keyword research and title ideas. The useful question is “which of my last 20 videos beat my average, and what do they have in common?” It’s much faster than clicking through the dashboard yourself.
8. ElevenLabs. Anyone who records anything. Turns text into natural speech, and turns recordings into text. Transcribing a meeting or a voice note and having the AI pull out the actions is the use I’d start with.
9. Composio. Anyone who wants a lot of apps without connecting each one. One connection that opens up hundreds of apps. It’s how I reach tools that don’t have their own connector. The trade-off is a middleman between you and the app, so check what it’s allowed to do.
10. A browser connector (Claude in Chrome). Everyone, for the long tail. When there’s no connector for a site, the AI drives a real browser like a person would: log in, read, click, type. Mine runs on a Mac mini that never sleeps. It’s slower and more fragile than a proper connector, but it covers what the first nine can’t.
Set it up this week
Five steps. Don’t skip the third.
- Pick one job that repeats. Inbox triage, the weekly sales report, the Friday update. Not “help me with everything”. One job.
- Connect one app. Not six. The app that job lives in.
- Read-only first. For the first week the AI can read and draft. It cannot send, post, delete or change anything. You find the mistakes while they’re still drafts.
- Write the instruction once and save it. In Claude, put it in a project’s instructions so every chat starts with it. Here’s the one I’d copy for inbox triage:
Every weekday morning, read my inbox from the last 24 hours. Sort it into three groups: needs a reply from me today, waiting on someone else, and noise. For the first group, draft replies in my voice (short, plain, no exclamation marks) and save them as drafts. Never send anything. Put anything that mentions money, a deadline or a complaint at the top. Keep the whole summary under 200 words.
- Run it by hand three mornings. If it’s right three times, schedule it. If it’s wrong, fix the instruction, not the output.
What it costs
The connectors themselves don’t cost extra on a paid Claude plan. The apps on the other end charge whatever they already charge you, and some of them charge by use. Image and video generation is the big one: a single short clip costs me about ten credits, and a plan can run dry faster than you expect. Check the pricing page before you connect, not after the invoice.
Where it actually goes wrong
These are all things that happened to me, not things I read about.
It scheduled something I hadn’t approved. An agent put a store video into the scheduler as live, when I had only meant it as a draft. Nothing went out because I caught it in time, but only just. The fix wasn’t a smarter AI. It was a rule: nothing goes out unless a cell in a spreadsheet says Approved.
The credits ran out and nothing broke loudly. I have two videos waiting to be redone because my balance is 1.49 credits and a clip costs 10. The AI doesn’t panic, it just reports it can’t do the job. The row sits there. Put a balance check at the start of any job that spends money.
One door was locked and the others weren’t. Instagram and TikTok posts go through fine. Facebook has been blocked for over two weeks by an identity check on the page. The scheduler happily says “scheduled” for all three. Check the live post on each platform, not the dashboard.
The login lapsed. One of my own posts missed its slot because the browser on my Mac mini had dropped its connection. The AI told me. I just wasn’t looking. Tokens expire, and the first sign is the AI suddenly “can’t find” something it found yesterday.
A five-legged dog went out. A store video that looked fine at a glance had a dog with an extra leg, and it was published. That one’s on me. An agent that can post is an agent that can post the wrong thing, so nothing public leaves without a checkpoint. Read, draft, approve, then act, in that order.
What I’d tell you
Typing into AI is the demo. Connecting it is the job. One connector, one repeating task, read-only for a week, three manual runs, then schedule it. Do that once and you’ll learn more about what these tools are for than a year of clever prompts will teach you.
If you want the build notes behind posts like this, The Build Log goes out by email and new subscribers get the Store Blueprint kit. Subscribe here.