12 Grok Bots for Meta Ads: The Full Setup, From Zero

The full setup, from zero

You commented "meta." Here's everything.

This is the actual system I run on our ad account and our clients' accounts. Twelve bots, split across research, audit, monitoring and optimisation. Every prompt is here. Every setup step is here. So are the parts most people get wrong, which is honestly the more useful half.

One promise up front: I'm not going to tell you this replaces a media buyer. It replaces the parts of a media buyer's week that are pure repetition — the checking, the pulling, the comparing, the noticing. That's a lot of the week. It isn't all of it.

Read the safety section. It's short and it will save you an ad account.

Three research bots that read your competitors' ads every morning and tell you what's working before you've opened your laptop.

Four audit bots that catch structural problems, tracking breaks and creative fatigue while they're still cheap to fix.

Three monitoring bots that report daily and stay silent otherwise.

Two optimisation bots that pause losers and step budget on winners, inside limits you set.

And one gap in the middle, which I'll be straight about: none of these bots make an ad. They find the angle. Something else has to turn it into a creative. More on that when we get there.

Five minutes here saves you hours later, because the architecture explains most of the weird behaviour you'll hit.

Grok Bot launched on 11 August 2026. It's still labelled early beta. It is not the Grok chat model and it is not a browser extension — it's a desktop and mobile app that gives you agents with their own cloud computer. They browse, run commands, read and write files, and sign into tools the way a person would.

Four things about that computer matter:

One computer, shared by all your bots. The machine is assigned to your account, not to each bot. All twelve of your bots share its filesystem, browser sessions, cookies and command-line credentials. xAI's own documentation says explicitly not to treat separate bots as a security boundary. If one bot logs into something, they all have it.

Each bot gets its own screen. They can work in parallel, but each bot runs one computer-use task at a time. Twelve bots doing browser work simultaneously is not the design.

Bots have separate memory. Each keeps its own conversation, role context and summaries of past work, even though they share files. This is what lets you give each bot a narrow job and have it stay in lane.

It reaches the internet from static IPs. Grok Bot computers use fixed egress addresses. Remember this — it comes back in the safety section.

The building blocks you'll use:

• Skills — reusable instructions for how to do a task
• Routines — instructions for when a bot runs, on a schedule or trigger
• Plugins — connectors to external services, including anything MCP

xAI's recommended sequence is: do it once as a task, get it reliable, save it as a skill, then automate it as a routine. That is exactly right and I'd add one thing — don't skip straight to the routine because the first run looked good. One good run is not reliability.

Plan and platform

Grok Bot runs on macOS (Apple silicon and Intel), Windows (x64 and Arm64), Linux (.deb, .rpm, AppImage) and iPhone on iOS 18+. There's no Android app and no iPad app.

Billing is where it gets confusing. Grok Bot runs on Cursor's infrastructure — it uses Cursor authentication, and your training opt-out follows Cursor's privacy settings rather than xAI's. Eligible plans in the documentation are SuperGrok Plus, SuperGrok Heavy, Cursor Pro+, Cursor Ultra, and Cursor Teams Standard and Premium. The marketing page lists a slightly different set including base tiers at lower prices. Check what your specific plan actually includes before you pay for a year.

If you hold both a Cursor and a SuperGrok plan, it draws from whichever has more usage remaining.

The cost reality nobody mentions

There are no published numeric usage allowances. What's documented from people running it: a six-agent setup consumed 42% of a week's quota on its first day. Roughly 100 basic completions plus a single ten-minute script came to about 5% of a week.

Twelve bots on daily routines will burn through a weekly allowance faster than you expect. Start with three. I'll tell you which three at the end.

What else you need

• A Meta Business account with admin access to the ad account
• A Meta system user token — ads_read for bots 4 through 10, ads_management for 11 and 12
• Somewhere to receive reports (email or Slack, both work as plugins)
• About two hours for initial setup, then twenty minutes per bot

Four rules. They're not optional and they're the difference between this being useful and this being expensive.

Rule 1: Never drive Ads Manager through the browser

Meta has a full Marketing API. Every bot in this guide that touches your ad account uses it.

Building campaigns, editing budgets or pausing ads by clicking through the Ads Manager interface with an agent risks account restriction. Ad account bans are slow and painful to reverse, and unlike an organic account you can't just start a new one. There is no upside — the API does everything the UI does, faster and deterministically.

The Ad Library is different. It's a public, unauthenticated page. Browsing it carries no account risk at all, which is exactly why bots 1 through 3 use the browser and every other bot uses the API.

Worth knowing: Meta's Ad Library API only covers political and social-issue ads. Commercial ads are visible in the web interface but not through the API. That's the specific reason an agent with a browser is genuinely valuable here rather than just a novelty.

Rule 2: Never automate posting from a logged-in social account

Instagram, TikTok, Facebook, LinkedIn and X all restrict automated account access in their terms. Combine that with static egress IPs and a daily schedule and you have about as fingerprintable a pattern as exists.

Research from public pages is fine. Publishing from your accounts is not. Keep publishing manual or use official APIs.

Rule 3: Set approval rules before the first run, not after

Grok Bot lets you set Require Approval and Always Allow rules. When they conflict, Require wins. The documentation warns against broad rules like "allow everything in the browser" — take that seriously.

For this system:

• Bots 1–10: Always Allow on read operations, browsing, file writes
• Bots 11–12: Require Approval on every write to the ad account, without exception

You can loosen bot 11 later once you trust it. Start tight.

Rule 4: Hand over credentials yourself

Grok Bot passes control back to you for passwords, passkeys, 2FA, CAPTCHAs and payment steps. Let it. Don't store account passwords in a bot's instructions or in a file on the shared computer — remember every bot can read that filesystem.

One more thing worth knowing. In June 2026, Adversa AI reported a context injection attack against Grok's web chat agent that could exfiltrate chat history and account details through a crafted URL; as of late August it was still unpatched. That's the chat product, not Grok Bot. But Grok Bot's public documentation doesn't discuss prompt injection at all, and an agent holding live logins is a much higher-stakes target than a chat window. Treat anything a bot reads from the open web as untrusted input — which is another argument for keeping write permissions narrow.

Install and sign in

Download the desktop app, sign in with your Cursor or SuperGrok credentials, and let it provision your cloud computer. First boot takes a few minutes.

Build the workspace

Everything shared lives in

. Create this structure first — the bots reference these paths constantly:

Connect the plugins

In the app, open Plugins and connect:

• Meta Marketing API (via MCP connector or direct API credentials)
• Email or Slack for report delivery
• Google Sheets or Drive if you want reports somewhere shareable

xAI's own docs say to prefer a connector over browser clicking whenever one exists. That's the right instinct and it maps exactly onto Rule 1.

This is the step people skip, and it's why their output is generic.

Create

before you build a single bot. Every bot reads it. Fill in every field.

That last block matters more than it looks. Without it you get walls of numbers. With it you get a report you can read in ninety seconds.

These use the browser against public pages. Zero account risk. This is where I'd start.

Bot 01 — Ad Library Scout

What it does: Sweeps every competitor's active ads daily. Reports what launched, what got killed, and how long each survived.

Why it matters: Longevity is the closest thing to a free performance signal you get. An ad running 60 days is surviving someone's optimisation decisions. An ad that lasted four days is their failed test — and their failed tests are free negative results for you.

Skill:

Routine: Daily, 6am.

What good output looks like: Ten to thirty rows per competitor, a short new/killed diff, and one or two flagged velocity spikes a week. If you're getting hundreds of rows, tighten the geo filter.

Bot 02 — Hook Miner

What it does: Transcribes the opening of every competitor video ad and clusters them into hook archetypes.

Why it matters: The first three seconds decide whether the rest of the ad gets watched. Everyone knows this. Almost nobody has a systematic library of what's actually working in their category right now.

Skill:

Routine: Weekly, Monday 7am.

Note: Weight by survival, not frequency. Counting ads tells you what competitors tried. Weighting by days running tells you what worked.

Bot 03 — Angle Gap Finder

What it does: Maps every angle running in your category, then shows you the ones nobody has touched.

Why it matters: Most people copy the crowded lanes because that's where the visible evidence is. The gaps are usually where the cheap attention is.

Skill:

Routine: Weekly, Monday 8am, after Bot 02.

That last instruction is the important one. This bot's output is a brief, not an ad. Which brings us to the gap.

Here's where the honest part of this guide lives.

You now have a system that reads your entire competitive landscape every morning, extracts what's working, and hands you a validated angle with a brief attached. That's genuinely the hard research half of media buying, running unattended.

And then it stops. Because a computer-use agent cannot make an ad.

It can browse, transcribe, cluster, compare and write. It cannot render a video that matches your brand, or produce a static that gets your typeface and spacing right, or generate a UGC-style talking head that doesn't look like a stock asset. That's a different class of problem — it needs generative models, brand extraction and a rendering pipeline, not a browser.

So most people who build this system arrive at a validated angle on Monday morning and then wait four days for creative. The research runs at machine speed and the production runs at human speed, and the whole loop moves at the speed of the slow half.

This is the exact gap we built HeyOz to close. You give it a product URL, it pulls your brand — typeface, colours, spacing, product imagery — and generates the ad. Video, static, UGC and carousel formats from the same input. So the brief that comes out of Bot 03 on Monday morning becomes a set of testable creatives the same day, in your actual brand rather than an approximation of it.

The workflow, end to end:

You can absolutely run the twelve bots without it and hand the brief to a designer or an editor. The system still works. It just runs at whatever pace your creative process runs at, and the research half will spend most of its week waiting.

Now back to the bots.

These read your ad account through the Marketing API. Read-only. Still no write permissions at this stage.

Bot 04 — Account Auditor

What it does: Walks your account structure looking for overlap, fragmentation, stranded budget and ad sets that never left learning.

Routine: Weekly, Monday 9am.

Bot 05 — Pixel Watchdog

What it does: Clicks your own live ads and verifies that events actually fire and UTM tokens actually resolve.

Why it matters: This is the failure that costs the most and gets noticed the latest. Dynamic UTM tokens arriving unresolved — literal

in the URL instead of the campaign name — silently breaks your attribution for however long it takes you to notice. I've had it happen. A bot checking daily catches it the same day.

Routine: Daily, 7am.

That last instruction is what makes a daily bot survivable. A report that only speaks when something is wrong gets read. A daily wall of green checkmarks gets filtered within a week.

Bot 06 — Fatigue Sentry

What it does: Tracks frequency and CTR decay per creative and flags the rollover before it happens.

Routine: Daily, 7:30am.

On the swap-in line: this is the second place the creative gap shows up. The bot tells you a winner is dying and you need a fresh variant of a proven angle. If your production pipeline takes a week, the ad dies before the replacement lands. If you can regenerate a variant of the same angle the same day, fatigue stops being a loss and becomes a scheduled refresh. That's the practical argument for having a fast creative engine sitting behind this bot.

Bot 07 — Leak Detector

What it does: Finds every ad set burning money above target with no path back.

Routine: Daily, 8am.

That separation is deliberate. The bot that finds problems and the bot that acts on them should be different bots with different permissions.

Bot 08 — Morning Digest

What it does: One report, every morning, before you open your laptop.

Routine: Daily, 7am.

The 200-word cap is the whole point. Reports that can't be read standing up don't get read.

Bot 09 — Anomaly Alarm

What it does: Stays completely silent until something actually breaks.

Routine: Every 4 hours.

An alarm that goes off when nothing is wrong is not an alarm. It's noise you'll mute in a fortnight and then miss the real thing.

Bot 10 — Test Scoreboard

What it does: Live standings on every creative test, with a decision rule and honest days-to-confidence.

Routine: Daily, 8:30am, while a test is live.

The tie rule matters. Small differences at low volume are noise, and a scoreboard that ranks them anyway trains you to make decisions on nothing.

These write to your live ad account. Approval gates on, no exceptions, until you've watched them for at least two weeks.

Bot 11 — Kill Switch

Routine: Daily, 9am. Approval required on every action.

The three-per-run cap is a circuit breaker. If your thresholds are misconfigured, you lose three creatives, not your whole account.

The kills log is quietly one of the most valuable files in this system. After a few months you have a record of every angle that failed and how much it cost to learn that — which feeds straight back into Bot 03's angle map.

Bot 12 — Scale Stepper

Routine: Every 3 days, 9:30am. Approval required on every action.

On the three-day cadence: you'll see advice saying to step winners 20% daily. Compounded daily that's roughly double in four days, and it will destabilise delivery and push the ad set back into learning. Every three days is slower and it holds.

Do not build twelve bots this weekend. You will hit the quota wall, everything will fail at once, and you won't know which part was the problem.

Week 1 — three bots. Bot 08 (Morning Digest), Bot 09 (Anomaly Alarm), Bot 01 (Ad Library Scout).

Read-only, immediately useful, and they teach you your burn rate before you've committed to anything. Run each one manually first. Check the output. Only then set the routine.

Week 2 — add research. Bots 02 and 03. Now you have the full research loop and you'll get your first angle map. This is the week you'll notice the creative gap, because you'll have a validated brief and nothing to render it with.

Week 3 — add audit. Bots 05, 06 and 07. Pixel Watchdog first — it's the one most likely to find something expensive on day one.

Week 4 — add the rest of monitoring. Bots 04 and 10.

Week 5 and beyond — optimisation, carefully. Bot 11 with approval gates, watched for two weeks before you consider loosening anything. Bot 12 after that.

By week six the whole system is running. If you try to get there by day three, you'll spend day four debugging twelve things simultaneously on a shared computer.

"Bot failed to respond" / bots stuck across devices. In late August 2026 a cluster of users hit exactly this — the shared cloud computer had entered a stuck state, confirmed by staff. The fix is resetting the bot's computer. Note that a reset can lose recent unsynced work, which is the argument for writing everything important to

files rather than leaving it in chat.

Quota gone by Wednesday. Reduce routine frequency before you reduce bot count. Most of these don't need to run daily — Bots 02, 03 and 04 are fine weekly. Anomaly Alarm every 4 hours is more valuable than most daily runs.

A bot doing something adjacent to its job. Tighten the skill and check the context file. Vague instructions produce improvisation.

Ad Library returning nothing. Check the country filter and confirm you're searching the Page name rather than the brand name — they're often different.

Bot pausing things you didn't expect. Check that Bot 11 is reading the current test definition. The most common cause is a kill rule firing on a creative that's mid-test and hasn't reached minimum sample.

Realistically, after a month: your competitive research runs itself, you find out about tracking breaks the day they happen instead of the week after, fatigue gets caught before it costs you, and nobody has to remember to check the account on a Sunday.

What it doesn't get you: strategy, taste, or a creative. It reads and it watches and it enforces rules you wrote. The judgement is still yours.

And it doesn't make ads. It hands you a validated angle and a brief and stops. If you have a fast way to turn that into a testable creative — HeyOz , a designer, an editor, whatever fits your setup — the loop closes and the whole thing compounds. If you don't, you'll have the best-researched backlog in your category and nothing running against it.

Build the three-bot version first. Get it reliable. Then decide how much further you want to go.

Questions on any of this — reply and ask. If you get one of these working and it finds something expensive, I'd genuinely like to hear about it.