ai-sales-team-claude is an open-source repo that turns Claude Code into a small sales team. Point it at a company URL and it researches the account, qualifies the lead with BANT and MEDDIC, maps the buying committee, drafts outreach, preps you for the meeting, and exports a PDF pipeline report, all from the command line. It is interesting to me because it is a public, generic version of exactly the pattern I have been building privately for Kaltura: multiple specialized agents doing the sales grunt work in parallel.
What it is
A set of Claude Code skills and subagents that automate the research-heavy front end of B2B selling. Built by Zubair Trabzada, MIT-licensed, Python 3.8+, and popular (about 881 stars, 299 forks at the time of writing). It packages 14 skills and 5 parallel agents behind a single /sales command router. It is one of a small suite from the same author, alongside ai-marketing-claude and ai-agency-claude.
Core commands:
| Command | What it does |
|---|---|
/sales prospect <url> |
Full account audit, 5 agents run in parallel |
/sales qualify <url> |
BANT + MEDDIC lead scoring |
/sales contacts <url> |
Decision-maker / buying-committee mapping |
/sales outreach <prospect> |
Cold, warm, and referral email sequences |
/sales prep <url> |
Meeting-preparation brief |
/sales report-pdf |
PDF pipeline report |
Plus more for follow-up, proposals, objection handling, ICP creation, and competitive analysis.
For business people
The problem it attacks is the hours a seller burns before the first real conversation: vetting a company, finding the decision-makers, scoring whether the deal is worth chasing, and writing the first email. This bundles all of that into a few CLI commands so one person gets the output of a small research team.
- Who it fits: solo founders doing their own prospecting, SDRs qualifying inbound, and agencies producing client-ready proposals and competitive one-pagers.
- Cost shape: the code is free (MIT). What you pay for is Claude Code usage; the heavier the research, the more tokens.
- The value: a repeatable, structured first pass. It grades each prospect 0-100 (A+ to D) from a weighted mix of firmographics, buyer fit, BANT, competitive signal, and outreach angle, so you triage where to spend real time.
- The honest limits: it researches from public signals only, no CRM, no proprietary data, no LinkedIn. Scores are estimates from what the web exposes, and the outreach is a first draft to edit, not to send. Treat it as an intelligent intern, not a closer.
For technical people
The architecture is a clean example of the orchestrator plus parallel-subagents pattern in Claude Code:
- A top-level
/salesorchestrator skill routes to 13 sub-skills (research, qualify, contacts, outreach, and so on). /sales prospectfans out to 5 agents that run simultaneously, each owning one dimension with a fixed weight in the composite score:
| Agent | Weight |
|---|---|
| Company research & firmographics | 25% |
| Decision-maker identification | 20% |
| Opportunity scoring (BANT) | 20% |
| Outreach strategy & messaging | 20% |
| Competitive intelligence | 15% |
- Qualification frameworks are coded in: BANT scored 0-100 across Budget, Authority, Need, Timeline; MEDDIC assessed across Metrics, Economic buyer, Decision criteria, Decision process, Pain, Champion.
- Stack: Python 3.8+, Claude Code required. Optional deps are lightweight: ReportLab (PDF), BeautifulSoup4 (HTML parsing), Requests (fetching). Layout is
skills/,agents/,scripts/(scraping, scoring, PDF),templates/(email and proposal playbooks). - The gotcha: the research layer leans on plain web scraping, which is fragile and rate-limited, and the "intelligence" is only as good as the public pages Claude can reach. There is no data moat here; the value is the orchestration and the scoring rubric, not any private source.
How to install
curl -fsSL https://raw.githubusercontent.com/zubair-trabzada/ai-sales-team-claude/main/install.sh | bash
Or clone the repo, run ./install.sh, and optionally pip install -r requirements.txt for the PDF and scraping extras. It runs inside Claude Code, so that has to be set up first. (As always, read an install.sh before piping it to bash.)
My take
This is worth studying, less as a tool I would adopt and more as a reference implementation of the pattern I already run in NicAI. The parallels are striking: an orchestrator command, specialized agents each scoring one dimension, a weighted composite grade, PDF output. Where it stops is exactly where a real system has to start, proprietary data. My own stack is wired into SQLite pipelines, a LinkedIn connections database, a Kaltura knowledge wiki, and real client files, so its research is grounded in what I actually know, not just what a scraper can find. This repo has none of that, and that is the honest ceiling of any generic "AI sales team": clean orchestration on top of public web data. The rubric is the reusable part. The scoring weights, the BANT/MEDDIC coding, and the parallel-audit structure are a good checklist to borrow, even if the plumbing underneath stays bespoke.
Further reading
