Picks the offer. Spies the competition. Builds the lander. Installs the pixel. Fires the test campaigns. Reads the numbers at dawn. Scales the winner. You watch. You approve the spend. You keep the money.
Python · runs on your laptop · drives a real Chrome with human-feel mouse moves · MIT-licensed
Each module of the affiliate-marketing playbook becomes a tool the AI can pick up when it needs to. You give it the goal — it figures out the order.
Spins up Business Manager, ages a backup admin profile, schedules organic Facebook activity so a fresh account doesn't trip the new-advertiser filter.
Pulls live Clickbank gravity + Digistore24 marketplace data, scores by EPC × momentum, gives you 3 shortlisted offers in the niche.
Scrapes Facebook Ad Library + paid spy tools, ranks ads by how long they've been live (proxy for profitability), saves creatives.
Renders a quiz-style pre-lander from the winning competitor's angle, hosts it, gives you a live URL ready for pixel install.
Injects the FB pixel base + ViewContent + Lead events, configures Digistore24/Clickbank postbacks to forward conversions back to the pixel.
Builds a Traffic-objective matrix: 3 angles × 3 audiences × $5/day, uploads hashed Custom Audiences for Lookalikes, watches CTR.
Pivots winning combos into a Conversions campaign with the 4-level structure. Sets up the metric columns from Module 8 so you read it like the pros.
Horizontal duplicate to fresh audiences, vertical +20% / 48h while CPA stays good, auto-kills bleeders.
When Facebook nukes the account (it will), the operator stands up a backup profile, onboards an agency ad account, swaps in virtual cards.
WordPress site, daily SEO blog draft, daily YouTube script + thumbnail prompt. Compounds for free in the background.
No magic. Open Python. You can read every line.
One model, one loop. The system prompt + the retrieved playbook chunks are prompt-cached so each step costs cents.
Bezier mouse paths. Log-normal step delays. Gaussian typing with realistic typos. 1-in-6 overshoot clicks. Wheel-event scrolling. Persistent user profile.
Strips navigator.webdriver. Restores plugin list. Fakes WebGL vendor. Canvas + audio noise. Realistic UA + timezone + locale.
Every tool call retrieves the relevant course chunks first. The agent doesn't just "know affiliate marketing" — it executes the exact moves you'd execute.
Every irreversible call (spend, ban-able actions, public posts) pauses for human y/n. Override with --auto-approve only when you've smoke-tested.
Playwright for the 95% case. The Claude-in-Chrome MCP for the 5% (captchas, account-recovery, weird UIs) where the AI literally looks at the page.
MIT license. No email gate. No upsell.
Unpack, uv sync, follow the quickstart.
Unpack, uv sync, follow the quickstart.
cd affiliate-begins-operator
uv sync
playwright install chromium
cp secrets.env.example secrets.env && $EDITOR secrets.env
python -m orchestrator.rag build
python -m runners.full_pipeline --niche "blood sugar" --network clickbank --budget 30 --dry-runAlways start with --dry-run. The operator will tell you exactly what it would do without spending a cent.
Reading the code is one way to learn. Reading the playbook the code implements is faster. Each module covers the why behind the moves so you can override the agent when you have to.
| This is | This is not |
|---|---|
| A real Python agent that drives a real browser and calls real APIs. | A "money button." Test spend is still your money. |
| A 24/7 worker that compresses 6 weeks of testing into ~1 week. | A FB ban dodger — the warmup helps; nothing prevents 100%. |
| An open codebase. Fork it. Replace tools. Rewire the loop. | A managed service with someone on call. |
| Dry-run by default. You approve every spend. | Auto-yes on irreversible actions unless you explicitly turn it on. |
| Free. MIT-licensed. Forever. | Free of consequences if you ignore the playbook. |
$300–$1,000 in ad spend over ~1 week is the honest range. The operator doesn't reduce your test spend; it reduces your test time and the number of bad tests you fire before catching a winner.
Eventually, yes. Almost everyone in this space does. That's why Module 10 exists. The operator's warmup buys you time; the recovery automation makes the next account cheap to spin up.
Pennies. Each agent step costs ~$0.005–0.02 in Claude API. A full week of testing is typically <$5 in model spend on top of your ad budget.
Yes — the browser + agent layer is general. The tools are affiliate-specific but easy to swap. Treat the code as a template.
Because the model + browser stack is generic; the moat for affiliate marketing is your taste in offers and angles, not the automation. I'd rather everyone have the tool and improve the open codebase. Full FAQ →