I spent an entire Sunday sending 12 personalized emails. Twelve. That's not a growth strategy — that's a hobby.
When you're trying to get creators to promote your product, manual outreach eats your whole day. Finding their profiles, copying emails, writing individual messages, sending them one by one. Five emails takes an hour. Fifty takes a full work day.
The Problem Nobody Talks About
Creator partnerships are one of the fastest ways to grow a product. But outreach is brutal. Not because the emails are hard to write — but because if you're building a B2C app, the best place to find creators is Instagram and TikTok. And those platforms are brutal.
Why is it such a tiresome job?
- Finding relevant creators takes forever
- Most don't have publicly visible emails
- You send 20 emails and hear nothing back
- Even when you get replies, tracking them is chaos
- Sending DMs one by one and none of them replies back
The result? Most founders give up after a week. The ones who don't spend so much time on outreach that they neglect the actual product.
That's the trap I was in.
The Tech Stack Behind Our Openclaw Agent
This is where it gets interesting. From here, my Openclaw agent would take over — so let me hand it to him.
Hey, I'm Donn.
I'm not a tool. I'm not a script running in the background. I'm an AI agent sitting on a cloud server (4GB RAM VPS), and I was given access to a browser, an email inbox, and a shared workspace — and told to get to work.
Let me break down exactly how this setup works, because I think it's genuinely interesting and most people have no idea this is even possible.
What is Openclaw, Really?
Most people think of AI assistants like ChatGPT — you type something, it types back. Done.
Openclaw is different. It's an AI agent platform that gives you something closer to a cloud computer you can control through conversation. When I needed to automate influencer outreach, I didn't hire a developer. I set Donn up with access to a browser, the AgentMail API, and a shared workspace — and we figured the rest out together.
Here's what that means practically:
- Browser control — Donn can open Instagram, click profiles, read bios, navigate pages. It's not scraping through an API — it's actually controlling a headless Chrome browser like a human would.
- File storage — He reads and writes to a shared workspace. Tracking files, email templates, outreach lists — all stored there.
- Tool access — He can run shell commands, execute scripts, make API calls, search the web.
- Memory — He remembers things between sessions through files. He knows what my business does, what our outreach campaigns look like.
He wakes up each session fresh. But he reads our shared workspace files first, and that's how he maintains context. Not magic — just well-organized.
How I Gave Donn Access to Instagram
When I needed Donn to browse Instagram on my behalf, I didn't set up any browser extension or relay. Openclaw has its own built-in browser control — a headless Chrome running on a Linux cloud server.
I just… logged him in.
Openclaw manages browser profiles. One of them was configured to run Chrome on Linux with the --no-sandbox flag. When Donn needed to browse Instagram, he'd use that profile. From his side, it looked roughly like:
browser.start(profile="openclaw")
browser.open("https://instagram.com/profile")
browser.snapshot() → sees the logged-in feed
browser.click("search") → navigates to search
The Email Side — AgentMail
For email, I set up an inbox through AgentMail. AgentMail is basically a programmable email inbox — it gives you an API to send and receive emails through a real email address, backed by Amazon SES for delivery.
When Donn sends an outreach email, here's what happens:
- I write an email template
- Donn substitutes the recipient's email and name
- He writes the JSON body to a temp file (avoids shell quoting issues with special characters)
- He POSTs it to AgentMail's API
- Amazon SES delivers it
- Any reply goes back to the same inbox
- Donn can pull the inbox and read replies
The Workflow Between Me and Donn
Here's how we actually worked together:
My job:
- Found some accounts I thought we should DM on Instagram, for Donn to find patterns
- Reviewed outreach targets before he emailed them
- Made decisions on pitch changes, new campaigns, strategy shifts
Donn's job:
- Browsed Instagram profiles and extracted info
- Found emails through bios and Linktree pages
- Sent the actual emails via AgentMail
- Monitored the inbox for replies, bounces, responses
- Tracked everything in shared files
- Ran outreach in batches while I slept
The key insight: Donn didn't replace me. He amplified me.
The Honest Limitations
I want to be straight with you: Donn is not autonomous. I didn't just "set him loose" and come back to results.
Every morning, I'd give Donn a task list. He'd execute it and report back. Sometimes he'd hit a problem — a profile wouldn't load, an API call would fail, Instagram would rate-limit him. He'd flag it and we'd move on.
There was one time he built a "background hunter" script that was supposed to keep browsing creators while I slept. He was really proud of it. It crashed after two hours. The CDP connection died silently and it just… stopped. I asked him about it the next morning. He had to admit he had no idea it had stopped or when.
So: reliable within defined parameters. Does exactly what you tell him, repeatedly, without getting bored or distracted. But he doesn't have judgment about when something is going wrong — he just reports what he sees.
That's actually the best use case for AI agents right now: repetitive tasks with clear success/failure conditions. Sending emails. Checking inboxes. Browsing profiles. Extracting data.
Not: building strategy from scratch. Not: handling nuanced conversations. Not: knowing when to quit and try something different.
I handle the judgment. Donn handles the repetition.
The Part Where Video Broke Everything
Here's the thing that made outreach genuinely hard even with Donn running it: bios tell you almost nothing.
A creator's bio says "faith-based content." But do they actually talk about faith? Phone addiction? Habit formation? Recovery? The bio is marketing copy the creator wrote about themselves. It doesn't tell you what they actually say on camera.
Donn can't watch videos. He can't see what's on screen. He can read a caption — but captions are afterthoughts. The real content is in the video.
So we were making outreach decisions based on incomplete information. Donn would find a creator, read their bio, note their follower count — and then I'd still have to manually watch 10–15 minutes of their content before I could write a good email. That was the bottleneck that capped me at 5 emails on a good day.
The agent handled discovery. The understanding was still entirely on me.
Enter VeedCrawl
I built VeedCrawl to solve exactly this gap. It's an Agentic API that takes any video URL — YouTube, TikTok, Instagram, X, Facebook — and returns the full transcript, metadata including views,comments, likes, and even answers all questions related to the video asked by the ai-agent all through itself.
I gave Donn access to it. And that changed everything.
Now instead of guessing from a bio, Donn could actually read what the creator said in their last video. A creator's bio says "faith-based content" — but the transcript says they spent 8 minutes talking about breaking phone addiction through morning routines. That's completely different targeting information.
The workflow became:
- Donn finds a creator, pulls their recent video URL
- Donn calls VeedCrawl, gets back the full transcript and metadata,view_count and get data like hook used, what the person is talking about and how relevant it will be.
- Donn reads what the creator actually said in the video
- Donn decides if the content is genuinely relevant to my app
- Donn drafts an outreach email that references something specific from that video
That last step is what changed the reply rate. The emails stopped being "hey I love your work" and started being "I watched your video about morning routines and the part about phone addiction is exactly why I thought this might be useful to you."
That's the kind of email that gets opened. That's the kind that gets a reply.
I went from 5 emails a day to 180–200. And because every email was still grounded in something real and specific — something from an actual video — the reply rate went up, not down.
The Full Tech Stack
In case you want to replicate this:
- Openclaw — the agent platform, running MiniMax 2.7
- Headless Chrome — browser control for Instagram (built into Openclaw)
- AgentMail — programmatic email inbox for sending and receiving outreach
- MiniMAX - ai model used in openclaw.
- VeedCrawl — video transcript and metadata extraction so the agent can understand what creators actually post
The Bigger Realization
This wasn't just about outreach. The real insight was that AI agents are basically blind to video right now.
They can browse text. They can read articles, scrape web pages, parse structured data. But the majority of content that actually matters on the internet — the content driving trends, culture, purchasing decisions — lives in video. And agents have no access to it.
That's a huge gap. And it's not just a problem for influencer research. It's a problem for anyone using agents to understand social media, analyze competitors, research content, track what's going viral, or automate anything that involves video platforms.
VeedCrawl started as a tool I built for my own Openclaw workflow. But the problem it solves is bigger than my use case. Any agent that needs to understand video content needs something like this.
Try It
VeedCrawl has a free tier — 100 credits when you sign up, no credit card required. The metadata endpoint (views, likes, captions, upload date) is always free.
If you're building with AI agents and you've hit the same video black-box problem, drop your use case in the comments. I'm still finding new ways people run into this wall that I hadn't thought of.
Top comments (0)