ethanplusai/harvey

The problem

Solo founders and small agencies that sell to local businesses need a steady pipeline of cold outreach, but the work is a loop: find prospects, research each one, write something specific, send, triage replies, book meetings. Existing sales-engagement platforms assume a sales team and a monthly tooling budget, and they are weakest exactly where small local businesses live. Harvey wraps that whole loop into one local Python process that wakes on a timer, decides what to do next, and keeps a dashboard of what it did.

What's different

The headline design choice is billing: Harvey drives the claude or codex CLI using your existing consumer subscription rather than a model API key, so per-token cost turns into subscription usage limits, which it attempts to read and throttle against. That is a real trade-off, not a free lunch — you are relying on provider terms and on the CLI's isolation flags, and the README says it pauses if the CLI is too old to isolate prompts. Second, prospecting is gated on a 29-signal catalog you explicitly confirm one by one, with costs shown, instead of a black-box ICP. Third, the decision loop is deterministic, derived from queue counts, so it does not spend a model call deciding what to do. Compared with Apollo/ZoomInfo-style data platforms, coverage of small businesses is the claimed strength; compared with sales-engagement suites, the difference is that it runs locally with no per-seat bill.

Who it's for

Best fit: a technical solo founder or two-person agency selling a service to local SMBs (roofers, dentists, HVAC), comfortable running Python, a mailbox on a secondary domain, and reading logs. Also reasonable for anyone who wants to read a working example of an agent with human approval gates. Look elsewhere if you sell enterprise, need CRM-grade reporting and team workflows, or are not willing to own deliverability and compliance yourself — the README itself points to legal and deliverability caveats rather than solving them.

Cost of adoption

Python 3.11+ (the README warns macOS system Python is too old), a logged-in Claude Pro/Max or ChatGPT plan, a sending mailbox (Gmail/Workspace suggested, ~$7/mo), and a dedicated secondary domain. Setup is a clone plus pip install -e . and an env file, or hand it to Claude Code via the shipped CLAUDE.md. Risks: it depends on undocumented-ish CLI behavior and --safe-mode, so provider changes can break it; subscription quota is shared with your own work; email deliverability and legal compliance are on you. MIT licensed. The README is unusually detailed, but there is no release or versioning signal in the input.

Verdict

The repository is about seven months old, and the only momentum signal available is a weekly Python board appearance at #12 with +1,283 stars this week — a spike, not a long track record, and it has never appeared on the daily board. Nothing on Hacker News suggests independent scrutiny. The idea is genuinely differentiated in its cost model and its confirm-each-signal stance, and the README shows real engineering care around isolation and budgeting. But the core dependency is someone else's CLI and subscription terms, which is exactly the kind of foundation that can change under you. Worth a weekend if you sell to local SMBs and enjoy operating your own tools; not something to build a business process on without a fallback path.