Open-Source Sellbird Runs AI Sales Outreach on a $7 Subscription
Sellbird, an open-source AI sales agent that runs on a Claude or ChatGPT subscription you already pay for, is climbing GitHub's trending list with 6.4k stars.
Takeaways
- Sellbird, an open-source AI sales agent, is trending on GitHub with about 6.4k stars and 358 new this week.
- It finds companies and decision makers, verifies emails, writes personal cold email and handles replies automatically.
- Model calls go through a Claude Pro/Max or ChatGPT Codex subscription, so there are no API keys or per-token bills.
- Plain code handles scheduling, signal detection and pre-send checks. The model only writes emails and drafts replies.
- Drafts need approval by default, and Claude calls run sandboxed with tools disabled to resist prompt injection.
- The license is AGPL-3.0 (formerly MIT as Harvey), and a hosted version at sellbird.ai is in early access.
Sellbird, an open-source agent built by developer ethanplusai, finds potential customers from a company website, identifies decision-makers, verifies email addresses, drafts personalized messages, sends approved campaigns, and handles replies. At publication, the repository had roughly 5,000 stars and 1,100 forks. Those figures show strong repository attention but provide no evidence of production adoption.
Sellbird’s distinguishing technical choice is its billing model. It invokes Claude through the Claude Code command-line tool on a Claude Pro or Max plan, or Codex through an eligible ChatGPT plan. This configuration requires an authenticated command-line session rather than a separate model API key, moving the workload from per-token billing to a subscription quota.
Costs move from tokens to quotas
Autonomous agents can accumulate API charges because they repeatedly inspect data, choose actions, and generate text. Sellbird reduces those calls by assigning deterministic work to regular Python code and reserving the model for tasks that require judgment. Its design notes publish the following estimates:
| Task | Estimated cost | Method |
|---|---|---|
| Business discovery | $0 with OpenStreetMap, or about $0.37 per 1,000 businesses through the cheapest listed paid source | Location and business-data queries |
| Website screening | $0 in model charges | Three ordinary web requests per company |
| Writing and sending | An existing AI subscription plus a mailbox | Claude Code or Codex for writing; an email account for delivery |
| Prospecting databases | $0 | No Apollo, ZoomInfo, Clearbit, or Clay integration |
The documentation uses about $7 per month for Google Workspace as its mailbox example and says comparable sales tools start at $250 to $500 per month. It also claims that a single commercial data provider typically covers about 30% of local businesses with five to 50 employees. These figures are project estimates, and prices, quotas, and data coverage can change.
A fixed queue controls the model
Sellbird assigns predictable tasks such as filtering records, checking database counts, and validating drafts to ordinary code. A scheduled “heartbeat” runs every 15 minutes, checks the remaining subscription quota, selects the highest-priority available job, records the result, and returns to sleep.
A fixed priority queue determines which work runs during each cycle:
- Handle replies.
- Send approved campaigns.
- Write campaign drafts.
- Prospect for new companies.
- Analyze results while idle.
Code-based heuristics detect buying signals, including missing online booking or an open role relevant to the seller’s product. The same layer filters low-quality results and runs preflight checks before delivery. Claude or Codex handles email writing and reply drafting, while a quota monitor slows the agent as usage approaches the plan limit.
Because scraped pages and prospect replies can contain prompt-injection attempts, the Claude path runs each invocation in safe mode with tools disabled. It also uses an empty temporary directory and excludes personal settings, plugins, and Model Context Protocol servers. These controls reduce the model’s access to local files and connected services when it processes untrusted text.
From clone to reviewed draft
Running Sellbird locally requires Python 3.11 or newer and an authenticated Claude Code or Codex command-line tool. The project recommends a separate sending domain with a gradually established sending history, commonly called domain warmup, to reduce the risk of mail providers classifying campaigns as spam.
The assisted setup uses the repository’s CLAUDE.md file to guide Claude Code. After cloning the project, a developer can open Claude Code in the repository directory and ask it to “set up Sellbird for me.”
Manual commands
git clone https://github.com/ethanplusai/sellbird.git
cd sellbird
pip install -e .
sellbird train https://your-company.com
sellbird signals --confirm free
sellbird discover
sellbird dashboard
sellbird run
The train command uses the company website as the starting context, discover searches for prospects, dashboard opens the review interface, and run starts the recurring heartbeat.
Every draft enters an Outbox for approval by default. A deterministic preflight check blocks messages containing unfilled merge tags, banned phrases, extra links, or undeliverable addresses. Lead-only workflows can stop after verification and use sellbird export to produce a CSV.
The small-business niche defines the limits
Sellbird targets founders, agencies, and small teams selling to local companies such as roofers, dental practices, and HVAC contractors. Public company websites often expose useful buying signals in this market, while commercial contact databases can have sparse coverage.
- Contact discovery: Sellbird finds at most two people per company and searches only names published on the company’s own pages, which limits its usefulness for large enterprise accounts.
- Subscription quotas: Claude or ChatGPT plan limits cap the agent’s throughput. The project advises operators to confirm that their intended usage complies with the provider’s terms.
- Deliverability and compliance: Domain setup, warmup, sender reputation, and compliance with applicable email and privacy laws remain the operator’s responsibility. Sellbird includes an optional compliance check.
- AGPL licensing: The GNU Affero General Public License can require operators to provide corresponding source code when users interact over a network with a modified version. Teams embedding Sellbird in a commercial service should review those obligations.
Harvey becomes Sellbird
Version 0.4.0 renamed the project from Harvey because another AI company already uses that name. Existing installations migrate automatically. An early-access hosted service manages domains, inboxes, and warmup, then connects with Claude, ChatGPT, Codex, or Cursor through the Model Context Protocol, a standard for linking AI clients with external tools and data.
A reusable agent architecture
Sellbird packages a broader development pattern: a deterministic scheduler manages routine work, a subscription-backed model handles judgment, sandboxing limits exposure to untrusted input, and human approval gates external actions. Prompts, sales playbooks, and configuration live in editable Markdown and YAML files, allowing teams to change outreach behavior without modifying Python. The repository offers a concrete reference for building persistent agents whose model usage remains bounded by design.