Cold Outreach Without the SaaS Bill: Sellbird Runs the Whole Pipeline on Your Own Machine
You've probably looked at cold outreach tools and done the same math everyone does: another monthly subscription, another vendor holding your contact data, another platform that wants to send emails on your behalf before you've had a chance to read them. What if the whole pipeline just ran locally, on the AI plan you're already paying for? That's the premise behind Sellbird, an open-source outreach platform that describes itself as running on your Claude or ChatGPT plan.
What It Does
Sellbird is a Python 3.11+ pipeline (AGPL-3.0 licensed) that handles the full arc of cold outreach: finding companies, figuring out why they'd buy, finding the person who decides, checking their email address, writing a personal email, sending it, and handling replies. It runs on your own machine, and by default nothing goes out until you approve it.
Each stage is something you can run end to end or stop at. Discovery searches for businesses matching your customer, free on OpenStreetMap or through a paid source you pick (with the cost of every paid search shown before it runs). Then it reads each company's own website for buying signals you've confirmed, like missing online booking, running ads, or a role they're hiring for. Enrichment finds the owner or manager named on the company's pages (two people per company at most), and verification learns the company's email pattern and checks one address, marking it verified, risky, guess, or invalid. From there it writes a short sequence per person and sends from your Gmail, Microsoft 365, or SMTP mailbox within your daily limit and their office hours. Replies get read, answered with drafts, auto-replies set aside, and opt-outs honored immediately.
Only want the list? Stop after verification and run sellbird export for a sequencer-ready CSV.
Why It's Cool
It treats your AI subscription as the compute layer. Instead of selling you tokens or model access, Sellbird leans on the Claude or ChatGPT plan you already have. That's a genuinely different cost structure from the usual outreach SaaS model, and it means the project doesn't need to markup inference to make money.
The signal-reading step costs nothing. Reading each company's website for confirmed buying signals involves no model calls and no cost. That's a smart split: cheap deterministic work happens cheaply, and the expensive generative work is reserved for writing and replies where it actually earns its keep.
Approval is the default, not a setting. Nothing goes out until you approve it. For anyone who's ever winced at a batch of auto-sent emails, this is the right default, and it's baked into the design rather than bolted on.
The verification model is honest about uncertainty. Addresses get marked verified, risky, guess, or invalid, and only deliverable ones are ever sent to. Most tools give you a binary "valid" flag that quietly hides how much guessing happened. This one doesn't.
Cost transparency on paid sources. Every paid search shows its cost before it runs. It's a small thing, but it signals the project was built by someone who's been surprised by an API bill.
You can stop anywhere. The pipeline is a series of discrete stages, and the export command turns it into a plain CSV at any point. That's a nice hedge: you're not locked into the sending half if you just want a good prospect list. (Worth noting: the project was formerly called Harvey, and existing setups carry over.)
How to Try It
The README's install section was truncated in what I could see, so check the repository for the exact setup steps. What's clear from the docs:
- You'll need Python 3.11 or newer.
- Clone the repo from github.com/ethanplusai/sellbird and follow the install instructions there.
- Once configured, kick off a run with:
$ sellbird run
- Review what it finds in the dashboard (the Today screen, the Outbox for approving drafts, and the signals to confirm), then approve sending from your own mailbox.
If you'd rather not run it yourself, sellbird.ai is the hosted version with sending domains, inboxes, and warmup handled for you — it's in early access. There's also a dashboard demo in the repo with invented data (a fictional booking product selling to roofers and dental practices), which is a good way to see the workflow before committing.
Final Thoughts
Sellbird is best suited to developers and small teams who are comfortable running Python locally and who'd rather own their outreach data than rent a seat on someone else's platform. The gated-by-default sending and the stop-anywhere pipeline make it a reasonable tool even if you only use half of it. It's early, and the README is the source of truth for what actually works today — but the architecture choices here are sound, and the "use the AI plan you already have" approach is the kind of pragmatic thinking more tools could borrow.
Follow @githubprojects for more developer tools and open source projects.