Hey, it's Cindy 🌱 You commented BRIDGE, so here is my exact connector setup.
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:)
💡 The one idea underneath all of this. Claude speaks MCP, the Model Context Protocol, an open standard for letting an assistant use outside software instead of only talking about it (the MCP spec). A connector is one MCP server you have plugged in. Zapier runs an MCP server that sits in front of its whole app catalogue, so adding that one connector hands Claude a very large toolbox at once.
why this matters
An agent with no tools is a chatbot
The reason most "build an agent in one prompt" tutorials do not survive contact with real work is that they assume the connections already exist. Writing the instruction is the easy half. Giving Claude a hand that can reach into your inbox, your email platform and your calendar is the half that takes an afternoon, once, and then never again.
🧠 The prompt is the plan It describes what should happen. On its own it produces a very nicely worded description of work that nobody did.
🔌 The connector is the hand It is what turns "schedule that post" from a suggestion into a thing that exists in your account five seconds later.
🌱 One bridge beats ten Instead of hunting for an MCP server per app, one Zapier connector fronts a catalogue you can add to whenever you need something new.
💡 Setup is one-time You wire it once and every future prompt inherits it. The cost is front-loaded, the payoff compounds quietly.
💡 The honest number. Zapier's own page currently says its MCP connects AI tools to "9,000+ apps" with "30,000+ actions across 9,000+ apps" (zapier.com/mcp). You will see 8,000 quoted a lot online, including by me, because that is what the page said not long ago. Either way the point stands: this is the widest single bridge available to Claude right now.
before you build
The cost gate, read this first
This is the part that catches people out three weeks in, so it goes before the setup rather than in a footnote at the bottom.
💵 Zapier's free plan gives you 100 tasks a month. Not 100 conversations. Tasks (Zapier pricing).
➗ Each MCP tool call costs two tasks. Zapier states that "each successful tool call through your MCP server consumes two tasks from your Zapier plan. Failed tool calls do not consume tasks" (Zapier's MCP help doc). So free realistically means around fifty actions a month.
📈 One prompt can be several calls. "Read my newsletter stats, then draft a post, then schedule it" is three tool calls, which is six tasks, from one sentence. Budget by actions, not by prompts.
🧾 Paid starts at $19.99 a month billed annually. That is the Professional tier at 750 tasks a month, which is roughly 375 tool calls. Free is genuinely fine for testing and for one or two small weekly routines. It is not enough to run your business on.
🌱 On the Claude side there is no paywall to clear. Custom connectors work on every Claude plan including Free, though Free is capped at one custom connector in total, and on Team or Enterprise an owner has to add it at the organisation level before members can use it (Anthropic's connector docs). If you are on Free, spend your one connector slot on this, because it is the one that fronts everything else.
the setup
Wire the bridge, about ten minutes
No terminal, no config files, no developer account. Have a Zapier login and a Claude login open in two tabs and work through these in order.
🗝️ Sign in to Zapier and connect the apps you care about. Go to zapier.com/mcp and sign in. Any app you have already connected to Zapier in the past is carried over automatically, so if you have ever built a Zap, you are further along than you think.
🆕 Create the server. Open mcp.zapier.com and click + New MCP Server. In the MCP Client dropdown choose Claude, give the server a name you will recognise later such as "Claude bridge", and click Create MCP Server.
🧰 Add your actions, and only your actions. On the server's tools screen, add the specific actions you want Claude to be able to run, for example Gmail "create draft", LinkedIn "create share update", Google Calendar "create event". A new server does nothing until you add tools, which is a feature: Claude can only ever reach what you deliberately handed it.
🔗 Open the Connect tab and copy your URL. Your server has its own private address on the Connect tab, along with client-specific instructions for Claude. Treat that URL like a password, because anyone holding it can drive the actions you just enabled.
➕ Add it to Claude. In the Claude app go to Customize, then Connectors, then Add custom connector. Paste your Zapier server URL, name it, save, and complete the sign-in prompt if one appears. Zapier notes that on Claude you need to be the owner of your Claude organisation to add a new server.
💻 Or do it in Claude Code. Same server, one line in your terminal: claude mcp add --transport http zapier YOUR-SERVER-URL. Then type /mcp to check it connected. The full command reference is in Anthropic's MCP docs.
🔛 Turn it on inside the conversation. Start a new chat, open the connectors menu in that chat, and switch the Zapier connector on. This is the step almost everyone forgets, then concludes the whole thing is broken.
🧪 Test it before you trust it. Ask Claude to list every tool it can see through the Zapier connector. If the list matches the actions you added in step three, you are wired correctly. If it comes back empty, you either added no tools or the connector is off in this chat.
💡 Start with drafts, not sends. For the first week, add only actions that create something you have to approve: create draft, create event, add row. Do not add "send email" on day one. You are building trust in a system that can act on your behalf, and the cheapest way to do that is to make every early mistake recoverable.
pick well
The six actions I would add first
Resist adding fifty tools. A shorter list makes Claude pick the right tool more reliably, and it keeps your task spend predictable.
✉️ Gmail: create draft The highest value, lowest risk action there is. Claude writes, you read and press send. Nothing leaves without you.
📅 Google Calendar: create event Turns "book that in for Thursday afternoon" into a real block with a real title, instead of a note you lose.
💼 LinkedIn: create share update Zapier's LinkedIn integration includes create share update and create company update (Zapier's LinkedIn integrations).
📊 Google Sheets: create row Your cheapest database. Anything Claude finds, researches or decides can be logged where you can sort it later.
💬 Slack: send channel message Good for a scheduled digest to yourself. A private channel is a lovely place for Claude to leave you notes.
⏳ Delay by Zapier The scheduling piece. Pair it with a posting action when you want something to go out later rather than now (Delay by Zapier).
💡 Where native beats the bridge. If Claude already has a first-party connector for a tool, use that instead and save your Zapier tasks. Check Customize, then Connectors, before you go build a bridge to something already sitting in the directory.
copy these
The 5 prompts I run through it
Every one of these assumes the connector is on in the chat. Swap the bracketed parts for your own tools and numbers.
Start here, always. This is the prompt that saves you an hour of confused debugging.
1 · map what you can actually do
List every tool you can currently see through my Zapier connector, grouped by the app they belong to, and for each one tell me in plain language what it does and whether it reads data or changes something.
Then tell me honestly what I cannot do yet with this set. If there is an obvious gap for someone who publishes content weekly and runs an email list, name the specific Zapier action I should go and add to close it.
Finally, flag any tool in the list that could do something irreversible if you used it carelessly, so I can decide whether to keep it enabled or remove it.
The chain from the reel. Notice it stops for approval before anything is published.
2 · newsletter to LinkedIn, end to end
Look at my most recent newsletter through the connected tools, work out which section got the most engagement, and turn that section into a LinkedIn post.
Judge the winning section on click data if you can reach it, and on open-and-position reasoning if you cannot, and say clearly which basis you used so I know how much to trust it. Then rewrite that section as a standalone LinkedIn post of 120 to 200 words that makes sense to someone who never saw the newsletter, opens on the specific idea rather than a throat-clearing line, and ends with a question people can answer from experience.
Show me the draft first and wait. Do not create the post until I reply with the word POST, and when I do, create it through the LinkedIn action and tell me exactly what you sent.
The one that quietly gives you back a morning a week.
3 · inbox triage into drafts
Go through the emails in my inbox from the last 24 hours and sort them into three groups: needs a real reply from me, can be answered with a short standard response, and needs no reply at all.
For everything in the second group, create a Gmail draft in my voice: warm, direct, no filler openers, no "I hope this finds you well", and under 100 words unless the question genuinely needs more. Do not send any of them. Leave every one as a draft.
Then give me a single short list of the first group, ordered by who is waiting longest, with one line each on what they actually want. If anything looks time sensitive or like it involves money, put it at the top and say why.
Because a chain you cannot inspect is a chain you will stop trusting.
4 · the weekly log
Every time you complete a task for me through the Zapier connector this week, add a row to my Google Sheet called AI actions log with the date, the app, the exact action you took, a one-line description of why, and whether I approved it first.
At the end of the week, read that sheet back and give me a summary: how many actions you took, which apps they were spread across, which ones I approved versus which ones ran automatically, and roughly how many Zapier tasks that consumed at two tasks per call.
Then tell me which of those actions were genuinely useful and which were busywork I would be better off not automating. Be blunt about the second list.
And the one that turns a good week into a repeatable one.
5 · the Monday setup block
It is Monday morning. Using the connected tools, put my week together in one pass.
Read my calendar for the next seven days and tell me where my genuinely free blocks are, in my timezone, ignoring anything under 45 minutes because I cannot do deep work in a gap that size. Then take my three priorities for the week, which I will paste below, and create calendar events for each one in the blocks you found, naming them clearly enough that future me knows what the block is for.
Before you create anything, show me the proposed schedule as a simple list and wait for me to say yes. If two priorities compete for the same block, say which one you would drop and why rather than squeezing both in.
be realistic
What the bridge is not good at
Knowing the edges is what keeps this useful rather than frustrating.
⏱️ It is not a scheduler Claude acts when you ask. For "post this on Thursday at 9am", pair a posting action with Delay by Zapier, or keep scheduling in the tool that already does it well.
📉 Deep analytics stay shallow Zapier actions are built for moving records around, not for rich reporting. For real numbers, use a native connector or export a file and let Claude read that.
🧭 Too many tools confuses it With sixty actions enabled, Claude picks the wrong one more often. Curate the list and keep it small on purpose.
💸 Retries are not free A prompt that quietly fails and retries burns tasks. Watch your Zapier task counter in the first fortnight so the bill never surprises you.
make it stick
Turn your best chain into one command
Once a chain has worked three times, stop retyping it. A skill is a small folder with a SKILL.md file inside that teaches Claude a process once, then triggers itself whenever your request matches. Paste this to build one.
✨ build the repurpose skill
Build me a reusable Claude skill called newsletter-repurpose that runs my newsletter to LinkedIn chain through my connected Zapier tools, and use skill-creator to package it.
The description should say: use this when I ask to repurpose my newsletter, turn my email into a post, or write this week's LinkedIn post. The steps are: find my most recent newsletter, identify the highest engagement section and state the evidence used, rewrite it as a 120 to 200 word standalone LinkedIn post in my voice, then show me the draft.
Two hard rules in the SKILL.md. Never publish anything without an explicit approval word from me in the message, and always report which tool you used and what it did after every action so I have a trail. If a tool it needs is not connected, it should say so plainly and stop rather than inventing the result.
📦 Take the folder it hands you. skill-creator finishes with a folder containing a SKILL.md. Ask it to zip that folder so it is ready to upload.
➕ Install it. In the Claude app go to Customize, then Skills, then Add, then Upload a skill, and choose the zip. In Claude Code, drop the unzipped folder into ~/.claude/skills/. Skills work on every plan, Free included (Anthropic's skills guide).
🌱 Use it as a sentence. "Do this week's LinkedIn post" now runs the whole chain, connector and all, while you do something that actually needs you.
when it goes wrong
The four things that break
"I don't have access to that app" The connector is off in this chat, or you never added that specific action in Zapier. Check the chat toggle first, then your server's tool list.
It sees zero tools A brand new Zapier MCP server has no tools by default. Go back to mcp.zapier.com and add the actions you want before blaming Claude.
Actions fail silently Usually the underlying app connection in Zapier has expired. Reconnect the app in Zapier itself, not in Claude, then retry.
Tasks vanish faster than expected Every successful call is two tasks and one prompt can be several calls. Ask Claude to report each tool call it makes, then reconcile against your Zapier task counter.