---
title: "Harvey Cold Email Agent for Claude Code: Setup, Costs and Limits"
date: "2026-10-28"
video_id: "7701766152866860321"
tiktok_url: "https://www.tiktok.com/@jackroberts____/video/7701766152866860321"
creator: "@jackroberts____"
category: "Marketing"
tags: ["harvey", "claude code", "cold email", "lead generation", "ai agent"]
excerpt: "Harvey is a free, MIT-licensed cold email agent that runs on your Claude plan. Here is what it needs to send, how approval works and where plan limits apply."
---
Harvey is an open-source agent at github.com/ethanplusai/harvey that finds prospects, profiles their websites, writes cold emails and works replies toward a meeting, using your Claude Pro or Max login. A TikTok from @jackroberts____ calls it free and hands-off. The code is free under the MIT license, but sending needs a dedicated mailbox, a warmup service and your approval on every email by default, and Anthropic's terms describe subscription limits in terms of "ordinary, individual usage".

## What is Harvey?

The [Harvey README](https://github.com/ethanplusai/harvey) describes a set of agents driven by editable Markdown skill files:

- a writer that drafts each email around specifics of the prospect
- a sender
- a reply handler that works conversations toward a meeting
- an analyst

Before writing, Harvey discovers prospects from several providers, profiles their websites for facts such as ad pixels, incumbent agencies and open roles, identifies owners and verifies email addresses. When I fetched the page it showed 3.1k stars, 631 forks, 376 commits and release 0.2.1. The owner appears only as the account "ethanplusai".

Harvey is also the name of an unrelated legal AI company. Nothing in the README connects the two.

## Is Harvey free?

The code is free. Running it is not. The README lists:

- Python 3.11 or later
- a mailbox to send from (Gmail or Workspace, or Microsoft 365), ideally on a dedicated secondary domain
- a mailbox warmup service, with TrulyInbox recommended through affiliate links
- a Claude Pro or Max subscription with the Claude CLI logged in

The default prospect source, OpenStreetMap, is free. Optional discovery sources such as DataForSEO, Serper and Scavio need their own accounts, and Harvey shows cost estimates before any paid call. Expect to pay for a domain, a mailbox, warmup and possibly lead data.

## How to set up Harvey

The video says "Just tell it what you sell and who your ideal customer is." The README's setup has more steps:

1. Configure `.env` and `harvey.yaml`.
2. Run `harvey train` against your own website.
3. Confirm the "signals" that define a good prospect.
4. Discover prospects.
5. Start `harvey dashboard` and `harvey run`.

The README also offers a path where Claude Code runs a setup wizard inside the repo. The video shows an animated terminal line, "harvey train bookwell.te", which matches the command name but is not output from a real run.

## Does Harvey send emails and book calls on its own?

Not by default. The video says Harvey "reads the replies and pushes those people towards a booked call for you." The README puts every outgoing email in an approval queue until you approve it. Autopilot is optional, and even then some emails are held for review.

The README also sets compliance and deliverability requirements:

- an opt-out line and a postal address in each cold email, set in config before the first campaign
- a `do_not_contact` list and an optional Precleared check for GDPR and PECR style rules, with a note that this is not legal advice
- LinkedIn automation off by default, because it violates LinkedIn's terms
- a dedicated domain with SPF, DKIM and DMARC, and a warmup of 2 to 4 weeks

A fresh domain cannot send at volume for weeks. Neither the video nor the README reports a reply rate, conversion rate or booked-call count.

## Can Harvey run on your Claude plan?

The README says model calls run through your logged-in Claude CLI, with no API keys or per-token billing. It lists Codex on a ChatGPT plan as an opt-in alternative on macOS and Linux only, which the video does not mention.

Anthropic's [Claude Code legal and compliance page](https://code.claude.com/docs/en/legal-and-compliance) adds conditions:

- OAuth login is intended for subscription purchasers and designed for ordinary use of Claude Code and other native Anthropic apps.
- Developers building products or services on the Agent SDK should use API key authentication.
- Advertised Pro and Max limits assume "ordinary, individual usage" of Claude Code and the Agent SDK.
- Anthropic may enforce these restrictions without prior notice.
- An end user may sign in to the unmodified Claude Code binary with their own subscription.

A personal tool that drives your own CLI sits closer to the permitted case than a service that routes other people's requests through your login. An unattended loop sending hundreds of emails is still a heavier load than the page describes, and Anthropic does not classify this exact pattern. Expect to hit usage limits on large campaigns, and read the current terms first.

## Harvey vs. a lead-gen agency

The video shows a mock agency invoice with line items for "Lead research", "List building", "Outreach" and "Account manager", and a "Cancel retainer" button. The amounts are blurred, so no price comparison is made. Harvey covers the research and outreach labor. You still carry the deliverability setup, compliance work and message review that a retainer would usually include.

The video's frames are animations, a typed command and a mock invoice. None shows a real Harvey run, a sent email or a reply thread.

## Bottom line

Harvey is a real, MIT-licensed outreach agent with a sensible default: nothing sends until you approve it. Budget for a sending domain, a mailbox, 2 to 4 weeks of warmup and possibly paid lead data. Running it at volume on a Claude subscription may run into Anthropic's usage terms.

## Sources

- [Harvey repository README, GitHub (ethanplusai/harvey)](https://github.com/ethanplusai/harvey): license, requirements, approval queue, compliance defaults, setup commands and repository figures.
- [Legal and compliance, Claude Code documentation, Anthropic](https://code.claude.com/docs/en/legal-and-compliance): subscription login, API keys for Agent SDK products and the "ordinary, individual usage" wording.
- [TikTok post by @jackroberts____](https://www.tiktok.com/@jackroberts____/video/7701766152866860321): the video that made the claims.


## Video description and verification notes

**The file is a 36.35 s vertical clip (1080x1920, 9:16) in an MP4/MOV container (major brand isom). The video stream is HEVC Main at 29.97 fps (2997/100), 1089 frames, 594,708 bit/s. The audio is HE-AACv2 stereo at 44.1 kHz, 64,345 bit/s. The total bitrate is 665,755 bit/s and the file is 3,025,193 bytes. The metadata lists format id bytevc1_1080p_665737-1, and its fps field is null, so I took the frame rate from ffprobe. The post timestamp is epoch 1793207152, which is 2026-10-28 17:05:52 UTC and matches the upload_date of 20261028. The uploader handle is jackroberts____, the channel name is jackroberts____, and the user id is 7404454526047028257. The sound is credited as "original sound" by jackroberts____. The description runs about 650 characters and links github.com/ethanplusai/harvey, followed by the hashtags #claudecode #harvey #coldemail #leadgeneration #aiagent. Engagement was captured on 2026-10-29 at about 12:03 UTC: 291,600 views, 21,600 likes, 83 comments, 3,625 reposts and 19,157 saves. I read five frames: 1, 5, 10, 15 and 18, so the sample is thin and I describe only what those five show. Frame 1 (about 0 s) is a split screen. The top half shows an animated wooden crate stamped "CLAUDE CODE" under a lamp. The bottom half shows a man with light-red hair and a mustache, wearing a dark cap that reads "Glaido", a pale yellow T-shirt and a chain, speaking at a boom-arm microphone, with the burned-in caption "built a". Frame 5 (about 8 s) shows a terminal bar reading "$ harvey train bookwell.te" with a cursor, a pixel-art figure in a suit and sunglasses, and the caption "what you sell". Frame 10 (about 18 s) shows two animated envelopes, one with a house icon and one with a tooth icon and a pen, under the caption "It writes". Frame 15 (about 28 s) shows a mock invoice titled "Lead-gen agency, Invoice · November" with line items "Lead research", "List building", "Outreach" and "Account manager", blurred amounts, a "Paid" badge, a "Cancel retainer" button and a falling balance chart, with the caption "to find". Frame 18 (about 34 s) is a collage of unrelated tiles: a crossed-out "Google Antigravity", "CURSOR" and "CLAUDE CODE", an "OpenA" logo with the caption "OpenAI just did", a man in a dark jacket, and a laptop spreadsheet, with captions that do not match the transcript. No real Harvey run, GitHub page, sent email or reply thread appears in any frame I read. The spoken argument is that an open-source Claude Code agent called Harvey finds leads, reads their sites, writes specific cold emails, handles replies toward booked calls, and replaces an agency using the viewer's existing plan. The transcript has 17 segments (ids 0 to 16) and its last segment ends at 35.92 s, inside the 36.35 s audio, so I found no text past the end of the audio.**

## "It's called Harvey, it is free and open source": verified, with costs the video omits

At 4.8 s the video says "It's called Harvey, it is free and open source on GitHub right now." The repository at github.com/ethanplusai/harvey carries an MIT license. When I fetched the page it showed 3.1k stars, 631 forks, 376 commits, and release 0.2.1. The page I fetched showed no creation date or release date, so I cannot say how recently the project was built, and the video's opener "Somebody just built" is not something I can date. The repository owner appears only as the account name "ethanplusai", and I did not find a personal name on the page.

"Free" covers the code, not the operation. The README says you need Python 3.11 or later and a mailbox to send from, ideally Gmail or Workspace, or Microsoft 365, on a dedicated secondary domain. It also calls for a mailbox warmup service and recommends TrulyInbox, using affiliate links. The default prospect source, OpenStreetMap, is free, but optional discovery sources such as DataForSEO, Serper and Scavio need their own accounts, and the README says Harvey shows cost estimates before any paid call. A viewer who takes "free" literally will still pay for a domain, a mailbox, warmup and possibly lead data.

The name "Harvey" is also the name of an unrelated legal AI company. Nothing in the video or README ties this project to it, and I did not investigate that further.

## "Just tell it what you sell and who your ideal customer is": partly matches the setup flow

At 8.4 s the video says "Just tell it what you sell and who your ideal customer is." Frame 5 shows a terminal line, "harvey train bookwell.te", which is consistent with the README, where you run `harvey train` against your own website. The README then has you confirm "signals" that define a good prospect before discovery. So the setup is not a single prompt. You configure `.env` and `harvey.yaml`, train on your site, confirm signals, discover prospects, then start `harvey dashboard` and `harvey run`. The README also offers a path where Claude Code runs a setup wizard in the repo. The frame shows an animation of a typed command, not captured output from a real run, so it demonstrates the interface name and nothing about results.

## "Finds the leads, reads their websites and picks the best fit": described in the README, not demonstrated

At 11.6 s the video says "And it finds the leads, reads their websites and picks the best fit for you." The README describes prospect discovery from several providers, free website profiling for facts such as ad pixels, incumbent agencies and open roles, owner identification, and email verification. That matches the claim as a description of features. The video offers no evidence of quality. None of the five frames I read shows a lead list, a profile, or a score, and the laptop spreadsheet in frame 18 sits in a montage whose captions do not match the speech, so I do not treat it as Harvey output. The fit judgment depends on the signals you confirm, so results depend on your input as much as on the tool.

## "Writes each cold email around something specific" and "pushes those people towards a booked call for you": the default is human approval

At 18.4 s the video says "It doesn't just blast the same email to everybody" and at 22.8 s "Then it reads the replies and pushes those people towards a booked call for you." Frame 10 shows a house envelope and a tooth envelope, which illustrates per-industry personalization but contains no email text. The README describes writer, sender, reply-handler and analyst agents driven by editable Markdown skill files, and says conversations are worked toward a meeting. That supports the claim in outline.

The video leaves out that the README makes sending gated by default. Every outgoing email waits in an approval queue until you approve it, and autopilot is optional, with some emails still held for review even then. The README also requires an opt-out line and a postal address in each cold email, to be set in config before the first campaign, offers a `do_not_contact` list and an optional Precleared check for GDPR and PECR style rules, and states it is not legal advice. It says LinkedIn automation violates LinkedIn's terms and is off by default. It gives a deliverability setup of a dedicated domain, SPF, DKIM and DMARC, and a warmup of 2 to 4 weeks. The phrase "for you" suggests a hands-off pipeline, while the documented default has you reviewing messages and waiting weeks before a fresh domain can send at volume. No evidence in the video or README shows a conversion rate, reply rate or booked-call count, so I treat any outcome as Unverified.

## "This runs on the core plan that you already have": the transcript word is mangled, and the plan claim needs a caveat

At 29.5 s the transcript reads "this runs on the core plan that you already have." The description says "It runs on your Claude plan", and the opening segment also reads "code code agent", so I read "core plan" as Whisper's rendering of "Claude plan". The README matches the description: it asks for a Claude Pro or Max subscription with the Claude CLI logged in, says model calls need no API keys or per-token billing, and lists Codex on a ChatGPT plan as an opt-in alternative on macOS and Linux only. The video does not mention the Codex option.

Anthropic's Claude Code legal page treats this arrangement with more care than the video does. It says OAuth authentication is intended for purchasers of subscription plans and designed to support ordinary use of Claude Code and other native Anthropic applications. It says developers building products or services on the Agent SDK should use API key authentication, and that advertised Pro and Max limits assume ordinary, individual usage of Claude Code and the Agent SDK. It also says Anthropic may enforce these restrictions without prior notice. The same page says an end user may sign in to the unmodified Claude Code binary with their own subscription. A personal tool that drives your own logged-in CLI sits closer to that permitted case than a service that routes other people's requests through your login, but an unattended loop sending hundreds of emails is a heavier workload than the "ordinary, individual usage" the page describes. I could not find Anthropic text that classifies this specific pattern, so whether it stays within plan limits is Unverified, and my search found secondary reports from early 2026 about Anthropic restricting some third-party harnesses, which I did not verify at the source. A user who relies on the plan for this should expect to hit usage limits on large campaigns and should read the current terms.

The video's contrast with an agency is also only partly fair. Frame 15 shows a mock "Lead-gen agency" invoice with a "Cancel retainer" button, with the amounts blurred, so no price is claimed. Harvey replaces the lead research and outreach labor, but you still carry the deliverability, compliance and review work that an agency retainer would typically cover.

## Key Takeaways

- **Verified:** Harvey exists at github.com/ethanplusai/harvey under the MIT license, as an agent that discovers prospects, profiles their sites, writes emails, handles replies and works toward meetings, per the README.
- **Partial correction:** The video says "free" at 4.8 s. The code is free, but the README requires a sending mailbox on a dedicated domain, a warmup service for 2 to 4 weeks, and optionally paid discovery accounts.
- **Partial correction:** The video says at 22.8 s that Harvey pushes people "towards a booked call for you." The README makes an approval queue the default, so nothing sends until you approve it, and autopilot is opt-in.
- **Clarification:** The transcript's "core plan" is most likely "Claude plan", based on the description and the README requirement of a Claude Pro or Max login. The README also lists a Codex alternative the video omits.
- **Unverified:** Reply rates, booked calls, lead quality and the project's creation date are not shown in the video, in the frames I read, or on the repository page I fetched. Whether a long unattended run fits Anthropic's "ordinary, individual usage" wording for subscription plans is also unverified.
- **Context:** The frames I read show animated illustrations, a typed command, a mock invoice and an unrelated news-style collage, with no captured Harvey output, so the video demonstrates the pitch and does not demonstrate the tool working.

## Resources

- [Harvey repository README, GitHub (ethanplusai/harvey), fetched 2026-10-29](https://github.com/ethanplusai/harvey) establishes the MIT license, the Claude Pro or Max and optional Codex requirements, the mailbox and warmup needs, the approval-queue default, the compliance defaults, the setup commands, and the star, fork and release figures I cite.
- [Legal and compliance, Claude Code documentation, Anthropic, fetched 2026-10-29](https://code.claude.com/docs/en/legal-and-compliance) establishes the wording on OAuth subscription login, API keys for developers using the Agent SDK, the "ordinary, individual usage" assumption behind Pro and Max limits, and the permitted case of a user signing in to the unmodified binary with their own subscription.
- [TikTok post by @jackroberts____, 2026-10-28](https://www.tiktok.com/@jackroberts____/video/7701766152866860321) is the source video, whose metadata, transcript and five frames (1, 5, 10, 15, 18) I read from the staged artifact directory.
