← all free guides
guide
Loading the guide…
Jarvis is Claude Code connected to the accounts you already use, with a voice added on top. Nothing new to download or install. Once it's set up, it can pull your RevenueCat numbers, check ad spend in Meta, clear out easy emails on its own, hand harder tasks to a subagent, and run the whole check-in every morning without you asking.
🧩
This is a stack, not a script. Each step below connects one account and works on its own. Do all ten, or just the two or three that solve something you actually have. Wispr Flow alone is worth it even if you never touch the Meta or RevenueCat steps.
✨ What you end up with
Once a few of these are wired up, a morning check-in with Jarvis sounds like this:
🎙️
You: "Hey Jarvis, catch me up before I get out of bed."
Jarvis: "On it, one sec while I pull it together. Downloads are up this week: 3,120 new installs and $5,610 in revenue. Ad spend came in at $612 with a 2.1 return, so that campaign's still healthy. Inbox-wise, you got 21 support emails overnight. I answered 17 of those off the FAQ file, mostly refund and login questions. The other four didn't clearly match anything, so I left them flagged instead of guessing."
You: "What about the crash reports from yesterday?"
Jarvis: "Handled. You'd already told me any confirmed onboarding bug goes straight to Tom, so I passed it along last night. He patched it, ran the tests, and it's sitting in a PR waiting on your review."
You: "Anything I should actually worry about today?"
Jarvis: "Two things. Tom's PR is ready whenever you are, and the street-interview ad angle is still outperforming the others, so I'd keep feeding budget there before testing anything new. Want me to line up Scout on new content angles while you look at the PR?"
You: "Yeah, go ahead."
Jarvis: "Done. I'll have something from Scout by the time you're back at your desk."
Every line in that briefing maps to a step below: revenue from Step 5, ad spend from Step 6, the resolved emails from Steps 7 and 8, Tom from Step 9, and the fact that it ran on its own before you asked, from Step 10.
✅ Before you start
- A computer (Claude Code runs on Mac, Windows, and Linux)
- A Claude subscription, Pro or higher (the Chrome extension and some MCP connectors need a paid plan)
- Accounts for whichever pieces you want: RevenueCat, a Meta Business account, a Google/Gmail account
- 30–45 minutes if you're wiring up everything, less if you're picking a few steps
You don't need to know how to code. Every connector below is a login screen or a copy-pasted command, and Claude Code walks through the configuration with you the same way it would in a terminal session.
🔧 The stack at a glance
| Piece | What it gives Jarvis | Step |
|---|---|---|
| Claude Code | The agent everything else plugs into | 1 |
| Wispr Flow | Talk to Jarvis instead of typing | 2 |
| ElevenLabs / Cartesia | Jarvis talks back, in a voice you pick | 3 |
| Claude for Chrome | Jarvis can click around your browser for you | 4 |
| RevenueCat MCP | Live app revenue and subscription data | 5 |
| Meta Ads MCP |
🤖 Step 1 — Install Claude Code
Everything else in this guide is a connector into Claude Code, so this comes first.
curl -fsSL https://claude.ai/install.sh | bashThat's the official native installer. No Node.js or Docker required. Full instructions for every OS (including Windows) are in the Claude Code Quickstart.
💡
Bonus tip from the video: once Claude Code is running, you can hand it a reference screenshot of a dashboard or UI you like and ask it to build something close to that. That's how the app stats dashboard in the demo above got built, in about five minutes. Drop the image in the chat and describe what should be live versus static.
🎙️ Step 2 — Give Jarvis ears: Wispr Flow
Wispr Flow is a voice-to-text keyboard that works system-wide, turning your spoken words into clean, punctuated text in any app, including your Claude Code terminal or the Claude app.
- Download it for your platform (Mac, Windows, iOS, Android) at wisprflow.ai/downloads
- Sign in and grant it accessibility / overlay permissions when prompted
- Hold your shortcut key (or tap the mic on mobile) and talk. It transcribes straight into whatever's focused, including a Claude Code session
Full walkthrough: Wispr Flow Setup Guide
Do this one first if you're only doing one. It's what makes talking to Jarvis feel like talking to an assistant instead of typing prompts.
🇬🇧 Step 3 — Give Jarvis a voice (optional)
Wispr Flow gets your voice in. This step gets Jarvis's voice back out: text-to-speech, so it can actually answer you out loud, British-butler accent included if you want it.
Option A — ElevenLabs (fastest to a specific character voice)
- Go to elevenlabs.io and create an account
- Voices → My Voices → Add a new voice → Voice Design
- Describe the voice in one sentence (e.g. "Refined British butler, calm and dry, mid-40s, unhurried pacing") and generate
- Grab the API key from your account settings to wire it into your Jarvis script or automation
Docs: Voice Design · Voices overview
Option B — Cartesia (built for low-latency conversational agents)
Create an account at cartesia.ai, grab an API key, and pick a voice from their library. Cartesia's Sonic model is built specifically for fast, natural-sounding agent replies rather than narration. See
🌐 Step 4 — Let Jarvis use your browser: Claude for Chrome
This lets Claude click buttons, fill forms, and read pages in an actual Chrome tab on your behalf. It's how Jarvis can open a dashboard that doesn't have an API and read the numbers straight off the screen.
- Make sure you're on a paid Claude plan (the extension is in beta, paid plans only for now)
- Install it from the Chrome Web Store (published by Anthropic) and confirm the publisher field says Anthropic before adding it
- Sign in with your Claude account when prompted
⚠️
Verify the publisher. Search results and copycat listings for "Claude" extensions exist. The official one is published by Anthropic. Check that field on the store listing before clicking Add to Chrome.
More on using it from a Claude Code session specifically: Use Claude Code with Chrome.
💰 Step 5 — Connect your revenue data: RevenueCat MCP
This is what lets Jarvis answer "how's the app doing" with real numbers: downloads, revenue, subscriptions, pulled live from RevenueCat instead of you checking a dashboard yourself.
Recommended — plugin install (OAuth, no key to manage):
In a Claude Code session, run /plugin, add the RevenueCat/ai-toolkit marketplace, then install the RevenueCat plugin. It authenticates via OAuth and scopes access to your existing RevenueCat account permissions.
Or — direct CLI connect with an API key:
claude mcp add --transport http revenuecat https://mcp.revenuecat.ai/mcp --header "Authorization: Bearer YOUR_API_V2_SECRET_KEY"Get the API v2 secret key from your RevenueCat dashboard under Project Settings → API Keys.
Full setup docs (Claude Code, Claude Desktop, and Claude web variants): RevenueCat MCP Server Setup
🔒
Use a read-only key if Jarvis only needs to report numbers.
📣 Step 6 — Connect your ad accounts: Meta Ads MCP
With this connected, Jarvis can tell you spend and ROAS without you opening Ads Manager. Grant write access and it can pause ads or adjust budgets when you tell it to.
- You'll need admin access on a Meta Business Manager account. No developer app or API token required
- In Claude, go to Settings → Connectors → Add custom connector, and add the MCP server at
mcp.facebook.com/ads - Complete the Meta Business OAuth flow. You'll be asked to pick a scope tier: read-only, read/write, or read/write/financial
🚧
This is a Meta beta, rolling out gradually per ad account. If you connect and get a "not enabled" error, your account isn't in the rollout yet. There's no waitlist to jump, just try again later. General Meta developer docs live at developers.facebook.com/docs.
Start with read-only the first time. Once you trust what Jarvis reports back matches Ads Manager, you can reconnect with a higher scope if you want it making changes.
📧 Step 7 — Connect your inbox: Gmail MCP
This step is what let Jarvis in the demo say "I resolved 13 automatically." It reads incoming mail, matches it against known questions, and drafts or sends replies.
Simplest path — Claude's built-in connector:
In Claude (web, desktop, or Code), go to Settings → Connectors → Add custom connector, name it Gmail, and connect. You'll get a Google OAuth prompt to pick the account and approve access. In a Claude Code session you can also just type /mcp, find Gmail in the list, and click Connect.
More granular / self-managed path — Google's own Gmail MCP server:
Google also publishes a first-party Gmail MCP server (currently in Developer Preview) that you configure through a Google Cloud project for finer control over scopes. Docs: Configure the Gmail MCP server.
🔒
Read vs. send are different permissions. Both paths let you grant read-only access first (search, list, read threads) and add draft/send access later once you trust what it's doing. Start read-only.
📚 Step 8 — Teach Jarvis your FAQ
This is the difference between Jarvis forwarding you every email and actually resolving most of them itself. There's no tool to install here, just a file.
- In your project (or a folder Claude Code has access to), create a file like
SUPPORT_FAQ.md - Fill it with your real, recurring questions and the exact answer you'd give: refund policy, how to cancel, common bug workarounds, account issues
- Reference it explicitly in your prompt or a
CLAUDE.mdfile: "Before drafting a reply to a support email, check the SUPPORT FAQ file. If the question matches, use that answer. If it doesn't clearly match, draft nothing and flag it for me instead."
💡
The quality of this file is the whole trick. The creator in the demo reported this resolves roughly 90% of incoming questions once the FAQ file actually covers what people ask. It's worth spending real time writing it out rather than a quick bullet list.
🧑💻 Step 9 — Build out your team: custom subagents
Subagents are how Jarvis hands work off instead of doing everything itself. In the demo, Tom is a developer subagent that picks up an already-approved coding task and works on it independently.
Run /agents in a Claude Code session and let Claude generate one with you, or write the file yourself. A subagent is a Markdown file with a small YAML header defining its name, description, allowed tools, and system prompt.
---
name: tom
description: Use for approved backend and app development tasks. Implements, tests, and reports back. Does not make product decisions.
tools: Read, Edit, Write, Bash
---
You are Tom, the developer agent. You only work on tasks that have
already been approved — if scope is unclear, stop and ask rather
than guessing. Report back what changed and why, concisely.Save it to .claude/agents/tom.md (project-level) or ~/.claude/agents/tom.md (every project). Repeat for a research-focused "Scout," a support-focused agent, or whatever roles match your own workflow.
Full reference: Create custom subagents
⏰ Step 10 — Put the morning briefing on a schedule
This is what makes the briefing happen on its own, before you ask for it: the difference between a tool you use and an assistant that's already working.
In a Claude Code session, type /schedule and describe it conversationally. For example: "Every weekday at 8am, pull my RevenueCat numbers, ad spend, and unread support emails, and give me a short spoken-style briefing with a recommendation for what to tackle first."
Claude Code offers three ways to run this, compared in the docs:
- Routines: run on Anthropic's cloud, keep going even if your laptop is closed
- Desktop scheduled tasks: tied to your machine
/loop: session-scoped, good for testing a prompt before you commit to a schedule
Full comparison and cron syntax: Run prompts on a schedule
💡
Test with
/loop first. Run your briefing prompt a couple of times manually or on a short loop interval before locking it into a daily Routine. It's much easier to fix a prompt that's still in front of you than one running unattended at 8am.🔧 If something isn't working
| What you see | What it means | Fix |
|---|---|---|
| Jarvis can't see a connector's data | OAuth wasn't completed, or the wrong account was picked | Re-run the connector setup in Settings → Connectors, watch for the browser OAuth prompt |
| Meta Ads MCP says "not enabled" | Your ad account isn't in Meta's beta rollout yet | Nothing to do yet, just try again later |
| A subagent isn't being used | Its description doesn't clearly match the task, or the file's in the wrong folder | Check it's in .claude/agents/ (project) or ~/.claude/agents/ (user), sharpen the description |
| Scheduled briefing silently stops running |
📱 More AI guides & tips
I make guides, tips, and content about AI and tech — the tools worth using and how to actually set them up.
📱
Instagram — @haejuicyy
TikTok — @haejuicyy
If you found this helpful, follow for more! 😛
Free guide
Unlock this guide
Drop your email and it's yours — plus the newsletter when new guides land.