Tool Friday #19 — PhantomBuster: I Gave Claude the API Key. It Stopped Being a Tool.

PhantomBuster is sold as a no-code tool. Point, click, scrape, done. That’s the demo, and the demo is fine.

The most valuable thing about it is the part they bury three menus deep in the docs: it has a REST API.

This week I did the obvious thing nobody seems to post about. I gave the API key to Claude. And PhantomBuster stopped being a dashboard I click — it became an agent I talk to.

The result, in one number: in a single hands-off run it pulled 927 decision-makers out of 159 Spanish companies — staged over 38 launches, account untouched, while I did other work. (I targeted ~190 companies; 159 had matching people to hand back.) That list would have cost me a week of manual exporting, or an SDR’s salary. That’s the post.

TL;DR: I pointed Claude at PhantomBuster’s API and it built a 927-person prospect list from 159 companies in one unattended run, account intact. The killer feature isn’t no-code — it’s the API that’s included on every paid plan. Hand it to a model and the tool drives itself: launch, poll, fetch, repeat, staged so it never trips LinkedIn’s limits. As a UI it’s a 6; as a programmable agent it’s a 9. The automation makes you faster, not safer, so you still go slow on purpose. $69/mo to start. Worth it at real volume, skippable if you scrape one list a month.

What it actually does

Strip the marketing and PhantomBuster is a hosted runner for “Phantoms” — small automations that log into your LinkedIn session and do one boring thing well. The one I live in is Search Export: give it a Sales Navigator search URL, it hands back every profile that matches as structured data.

That’s it. It’s the muscle. It is not the brain. And for two years I treated it like most people do — open the dashboard, paste a URL, click launch, come back later, download a CSV. It works. It’s also a chore you have to remember to do, one search at a time.

Giving it to Claude

Here’s the line straight out of the harness README I wrote this week:

The “PB × Claude” integration is just API key + function calling — this is that, scripted and auditable.

No magic. PhantomBuster exposes a v2 REST API. Auth is one header (X-Phantombuster-Key-1). There are exactly three calls I care about: list your Phantoms, launch one, fetch a container’s result. That’s a tool a model can use the same way it uses anything else.

So I wrote thirty lines of bash and jq — no MCP server, no framework, just curl and a key — and now Claude drives it. I say “pull the marketing directors at a couple hundred companies” and it runs the whole loop: launch the Phantom, poll the container until it’s finished, pull the JSON, wait, do the next one. For the dream run that was 38 launches back to back, completely unattended.

The reason it’s bash + jq and not some heavyweight integration: I can read every line. When a scrape returns nothing, I can see the exact request that was sent. A black-box “AI agent” that scrapes LinkedIn is a great way to get surprised. An auditable one is a tool.

The part that’s actually hard

The automation is the easy half. Two things make or break it, and neither is PhantomBuster’s job.

The URL is the whole game. Feed it a sloppy “marketing managers in Spain” search and you get a sloppy list. The method I use builds the Sales Navigator URL on company IDs — LinkedIn’s internal numeric IDs — not names. The CURRENT_COMPANY filter matches the exact company, so there’s zero name-ambiguity, zero “wrong Acme Corp.” Exact companies, exact seniority, exact region, in one URL. That precision is mine to build. PhantomBuster just runs whatever I hand it.

Here’s the proof it matters: across all 927 people the run returned, there was exactly one duplicate. Build a search on company names and you’d be wading through “wrong Acme Corp” all day; build it on IDs and the dedup does itself. The clean list is a side effect of the targeting, not a cleanup step afterward.

Going slow is a feature, not a bug. This is the one everybody learns the hard way. The API lets you fire a hundred scrapes in parallel. Do that and LinkedIn flags the account behind it — the account that is your real professional identity. So the harness does the opposite of what “automation” implies: five companies per launch, one container at a time, a 45-second pause between each, reusing one saved session. It also skips any group it already scraped, so a crashed run resumes instead of restarting. Slower on purpose. The automation makes you faster at the boring parts; it does not make you safe from the platform. That part is still on you.

What I’d want a friend to tell me before wiring this up

Not the feature list — the things you only learn after.

The API is included, but the session is the risk. Every paid plan has the API; you’re not paying extra to automate. But the moment you connect a LinkedIn account, that account carries the whole exposure. Don’t wire your main profile into an aggressive loop and hope.

It won’t fix a bad list. If your targeting is vague, automation just produces garbage faster. Spend your time on the URL, not the tool.

The terminal is the price of admission. The “give it to Claude” version assumes you’re comfortable with an API key, a .env file, and reading a shell script. If that sentence made you tense, the UI still works — you’ll just hit the manual ceiling sooner.

Yield is lumpy, and that’s normal. In my run one group of five companies returned 108 profiles; two others came back with one and zero. That’s not the tool failing — it’s that some companies are wide open on Sales Navigator and some are nearly invisible. Don’t judge a run by its best group or panic at its worst. Judge the total.

Credits expire monthly. Execution hours, Phantom slots, email credits — all capped per plan, all reset at the end of the cycle. Unused time doesn’t roll over. Buy the tier that matches your real volume, not your ambitious volume.

Who should bother

Wire it up if: you’re running real net-new prospecting (hundreds of companies, not dozens), you already build precise Sales Navigator searches, and you’re comfortable handing an API key to a model and reading what it does with it. This is where PhantomBuster goes from “fine tool” to “quiet engine.”

Skip it if: you scrape one list a month, or the terminal isn’t your world. The manual export genuinely covers you, and you’ll save $69 and a setup afternoon.

The verdict

Score: 8/10 — a 6 as a dashboard, a 9 as an agent, and the gap between those two numbers is the entire point.

PhantomBuster’s real product was never the no-code interface. It’s the API sitting quietly behind it — the thing that lets a tool stop being something you operate and start being something a model operates for you. That shift, from “open the app and click” to “tell the agent and read the log,” is the most underrated move in my stack right now, and it cost one afternoon of scripting.

Would I pay $69/mo for the dashboard alone? Probably not. Would I pay it for an auditable scraping agent that runs itself, account-safe, while I do something else? Already am.

The lesson travels past this one tool: half the software you already pay for has an API you’ve never opened. That’s where the upside is hiding. PhantomBuster is at phantombuster.com. Go find the API key.

Tool Friday is a weekly series where I review one tool I actually use. This week: a no-code scraper that gets ten times more useful the moment you stop using it no-code — and the honest case for, and against, handing your tools to a model.