Skip to content

Autonomous AI sales agent powered by Claude Code or Codex. Finds prospects, writes cold emails, sends campaigns, handles replies, books meetings. No API keys needed — runs on your Claude or Codex subscriptions.

19.8 KB
Raw Blame

Changelog

Notable changes to Harvey. Dates are release dates; the format follows Keep a Changelog loosely and versions follow semver, with the caveat that Harvey is pre-1.0 and minor versions can still change behaviour.

[Unreleased]

Added

  • harvey warmup [domain]: checks whether a sending domain is ready for cold email. It covers MX, SPF (+all, duplicates, top-level lookups), DKIM, DMARC, domain age, domain and SMTP-IP blocklists (a list that refuses the query is reported as such, not as a listing), warmup progress and bounce rate, then gives a READY / NOT READY verdict. It's read-only, and a DNS timeout is reported as unknown, never as a missing record.
  • A cold-email ramp for new domains (channels.email.warmup). From start_date, no cold email goes out for min_days; then the cold cap starts at start_sends and climbs by daily_increase a day up to max_daily_sends, counted per local day. Gmail, Microsoft and SMTP only: Instantly has its own warmup and per-account limits. Replies are never held. Off unless start_date is set. Both based on #18 by Tariz Shahid.
  • A postal address and opt-out line under every cold email (CAN-SPAM). Set compliance.postal_address in harvey.yaml; compliance.opt_out_line defaults to Not interested? Reply "no thanks" and I won't email again. Both are added when the email sends (native providers and Instantly), never to replies, and the Outbox shows them under each draft; "Open in Gmail" includes them too. Without an address (or with the opt-out line turned off), cold email still sends, and harvey run, harvey status and the Outbox tab warn until it's set. The opt-out line must ask for a reply Harvey recognises and can't contain a link. Based on #13 by Camilo Cruz.

Fixed

  • A follow-up no longer goes out before, or right after, the email it follows. Each step now waits until the previous one has actually been sent and its gap (at least a day) has passed. Before, approving a batch of drafts late, or resuming after a pause or a compliance hold, could send a whole sequence within minutes, or step 2 before step 1. A follow-up whose earlier email was rejected or failed is cancelled instead of arriving as a first email.

  • Windows: text files (harvey.yaml, .env, prompts, skills, trained output, tokens, exports) are read and written as UTF-8 instead of the locale code page, so harvey train no longer crashes on characters like "★" and prompts aren't garbled; command output is UTF-8 too, so the dashboard's log of harvey run can't crash on "⚠". An npm install's claude.cmd shim is resolved to the native claude.exe it wraps and is never run through cmd.exe, which would re-parse Harvey's isolation flags; if the executable is missing, Harvey says how to fix it. Files an older Harvey saved in the Windows code page (trained skills, harvey.yaml) still load, with a warning to re-save them as UTF-8. harvey export CSVs are now UTF-8 on Windows too, which sequencers expect (Excel needs Data > From Text/CSV to show accents). Based on #18 by Tariz Shahid.

  • Replies are read without the email they quote. The model, the stored thread and the answer get the reply without the quoted email, and any line that repeats what Harvey sent that person is set aside, so Harvey's own words (its footer, or an escalation word like "attorney" in its pitch) don't count as theirs. Opt-out and escalation words the person wrote are still found anywhere, including between or under quoted lines. "No thanks" / "no thank you" as a phrase, and a reply of just "STOP", now count as opt-outs.

  • The Sender claims each email right before sending it. A reply, bounce, hand send or rejection that cancels an approved email while the Sender is working through its queue now stops it, instead of the Sender sending its earlier copy and recording it as sent. Two Harvey processes running at once can no longer send the same email twice. An email cut off mid-send (Harvey stopped, or the mail library crashed mid-call) is marked failed and never retried, since it may have gone out; it shows under "Didn't send" with a note to check your Sent folder, and the dashboard won't offer to email that person by hand. Rejecting an email that is already sending or sent is refused instead of relabelling it.

  • Keep local website requests out of private networks, including redirects and DNS rebinding. Preserve the original hostname for TLS verification. Website fetching (search, company sites, profiling, the training crawler) now ignores HTTP_PROXY/HTTPS_PROXY, so it can't be routed around that boundary; mail and API integrations are unchanged. harvey train refuses local or intranet URLs and says so.

  • Instantly users: check require_approval. Instantly has no per-email approval queue, so with channels.email.require_approval: true (the default) Harvey now holds campaign deployment and hands replies to a human instead of silently bypassing the setting. harvey run and harvey status say when this is happening; prospecting and drafting carry on (drafts wait, using model quota, until sending is allowed). Set require_approval: false for Instantly autopilot.

  • Disable LinkedIn by default in the model, sample config, and product trainer; save the setup wizard's explicit LinkedIn choice when updating a config. Answering "no" to LinkedIn when re-running harvey setup now turns it off (saved credentials stay in .env).

  • Protect the local dashboard with an access token, restrict requests to trusted local origins, and save settings with owner-only permissions. Update and restart Harvey, then open the dashboard link printed in your terminal.

  • Isolate Claude calls from personal instructions and hooks, and fix Claude CLI installation in Docker. Thanks to @EICCIO-Advisors for the private report.

Added

  • Send by hand from the dashboard, and a live status panel. With no mailbox connected (or email turned off), each campaign's first email has an Open in Gmail button that fills in a compose window in the account Harvey sends as, and a Mark as sent box. Harvey never sends to someone marked sent by hand, even if a mailbox is connected later, and only offers it for people it could still email (not opted out, replied, bounced, or already queued by Harvey). The Today tab gains a panel showing each pipeline step, what's running, and what to do next.

  • Outbox review: see what a draft is based on, and edit it. Each pending draft now shows its campaign, the company (name, clickable website, description), the contact and how they were found, email status and score, tech / agency / hiring / ads / booking signals, and the Writer's notes, so every claim can be checked against what Harvey actually found. Edit (or E) turns subject and body into fields; Save rewrites the draft only while it is still pending_review and never approves it. An edit the sender would block (merge tags, HTML, banned phrases) is refused on save with the reason; a reply keeps its subject so it stays in the thread; every edit is logged with what it replaced. While editing, the shortcuts and Approve all can't act on the unedited draft. POST /api/outbox/{id}/edit backs it.

  • Microsoft 365 mail provider. Set channels.email.provider: microsoft to send and read through a Microsoft 365 / Outlook mailbox via Microsoft Graph, for domains whose mail lives in Exchange Online (which no longer accepts passwords over IMAP). Signs in once with OAuth + PKCE (harvey microsoft auth; no client secret, no stored password), drafts then sends so the Message-ID is known, threads replies with createReply, reads the last 7 days of the inbox without marking anything read, refuses to act as any mailbox other than persona.email, and holds sending when the inbox can't be read. See docs/microsoft365.md.

  • Optional per-recipient compliance check (Precleared). Set compliance.precleared.enabled: true with sender_country: GB or US to have each new recipient checked before the first cold email against UK PECR, US CAN-SPAM, Canada's CASL, Germany's UWG and France's rules. Any verdict but ok_to_email fails that outbox item with the rule and the next step. If the check can't be made (daily allowance used up, network down, a malformed response), cold emails wait for a later cycle while replies still go out. One check per address, cached for 30 days; changing the sender country, address source or product description triggers a fresh check. The receipt id is logged with each send. Native providers only. Off by default.

  • Cohort builder pagination and CSV export. The cohort table on the Signals tab used to stop at the first 50 matches with no way to see the rest. It now pages with Prev / Next ("Showing 51–100 of 147"), and an Export CSV button downloads the whole cohort. POST /api/cohort accepts offset and limit (default 50, max 200) and orders by name so pages never overlap; GET /api/cohort.csv?require=...&exclude=... returns every match. Counts and exports include every match; pages use a stable name/id SQL order. CSV downloads carry dashboard authorization and escape spreadsheet formulas.

[0.3.0] — 2026-10-13

Added

  • Opt-in Codex brain (ChatGPT plan). Set brain.provider: codex (or pick it in harvey setup, or harvey train --provider codex) to think with the Codex CLI on a ChatGPT subscription instead of Claude. Claude remains the default and there is no automatic fallback between providers. Each call runs a throwaway, isolated codex app-server with no tools; Harvey verifies the lockdown from Codex's own reports before sending any prompt, aborts on any attempted tool use, refuses API-key logins and custom endpoints (checked before your login is exposed to Codex), and never changes your own Codex settings or login: each call gets a private copy of the login without its refresh token, so an expired login must be renewed with Codex itself or codex login. Requires Codex CLI 0.159.3; other versions are refused. See the README section "Using a ChatGPT plan instead (Codex)".
  • brain.model picks the model for either provider, for Harvey only. For Codex use gpt-5.5 (suggested by harvey setup): the verified text-only model. Blank auto-selection only picks listed models without code-execution tools, and on current ChatGPT accounts there are none; code-mode models are refused.
  • If Codex can't be used (expired or renewal-due login, unsupported CLI, unsafe Codex config, rejected model, plan usage limit), Harvey pauses the heartbeat instead of running agents that would fail: replies stay unread and unescalated, queued emails wait, and the reason shows in the log, harvey status and the dashboard. It resumes on its own once the cause is fixed. harvey status --check runs the same check (no prompt sent) on demand and exits non-zero when it fails. The Codex provider supports macOS and Linux; it is refused on Windows (use Claude there).

Changed

  • Each heartbeat now sends email after the agents that read and answer replies have finished, and native mail providers (Gmail/SMTP) read the mailbox every cycle, so a reply always cancels its queued follow-ups before that cycle's sends (previously the two could run at the same time, and a reply arriving with no open conversation waited for a later cycle). This applies to Claude and Codex alike.
  • A reply is only recorded once its reply draft (if it needs one) is written, so a brain outage part-way never half-records it. If a working brain returns no usable draft, the conversation is flagged for a human instead of being left open with no answer.
  • Usage is tracked per provider. The budget gate, harvey usage, harvey status and the dashboard read and label the active provider's own quota; existing usage history is kept and attributed to Claude.

Fixed

  • Dashboard HTML and text assets now use UTF-8 explicitly, fixing Windows decoding failures. Regression tests cover HTML, CSS, JavaScript and SVG.

[0.2.1] — 2026-10-13

A security release. Update if you run Harvey.

Security

  • Claude now runs with no tools. Every Claude call used to run with --dangerously-skip-permissions, and prompts include text Harvey doesn't control: prospect email replies and scraped web pages. Instructions hidden in that text could make Claude run commands, read or write files, or use the operator's connected MCP servers. Calls now run with --tools "" and --strict-mcp-config; Harvey only ever needed text back, so nothing else changes. Reported privately by @hackchang. If you ran an earlier version with live replies coming in, check the account that ran Harvey for unexpected files or activity.

Fixed

  • harvey run no longer drops into the setup wizard when the trained config lives in harvey.local.yaml; the dashboard reads it too.
  • A heartbeat cycle is capped at 10 minutes, so one hung agent can't freeze Harvey, and quiet hours are re-checked every 5 minutes so a Mac waking from sleep resumes on time.
  • Sites that only answer on plain HTTP or www. are no longer marked unreachable.
  • Profile and contact backfills rotate through every company instead of retrying the same few forever.
  • The trainer writes full place names ("Solon, Ohio") and real job titles, so discovery doesn't geocode to the wrong city and found contacts aren't silently discarded.

Added

  • Contacts are backfilled for companies found without any, from their own team pages and, if a key is set, Hunter's named-people search.
  • HUNTER_MONTHLY_SEARCHES (default 25) sets one Hunter budget shared by every domain search.

Thanks to @sshaw0729-dotcom for the fixes above (#2).

[0.2.0] — 2026-09-13

The release that made Harvey worth running. The previous version could write and send email but could not reliably find anyone to send it to, and it needed a paid third-party platform to send at all. Both are fixed.

Added

  • A prospecting pipeline that works. Four stages, each independently runnable and re-runnable: DISCOVER (who exists) → PROFILE (what they are) → ENRICH (who decides) → VERIFY (can you reach them).
  • DISCOVER, with a provider menu. Four sources behind one adapter interface, shown side by side in the dashboard with what each does, what it costs, its free tier and which keys it needs. The default — OpenStreetMap — needs no account, no card and no key, so the whole pipeline can be run before paying anyone. harvey discover.
  • Cost discipline in the collector. Every discovery run estimates before it spends, refuses to start over your cap, re-checks the cap between batches, reads a kill switch, and flushes observations per batch so a run that dies keeps what it learned. harvey discover --estimate prints the projection and exits.
  • PROFILE. Free — three HTTP requests per business, no browser, no model call. Reads the incumbent agency out of the site footer, ad pixels, missing schema, abandoned blogs, open roles and named people. Chained onto DISCOVER and run on every heartbeat. harvey profile.
  • A signal catalog you confirm. Harvey knows how to collect 23 signals and collects none of them until a human says so. Each shows a plain-language description, what it costs, and how many companies already carry it. harvey signals.
  • The cohort builder. Pick what a prospect must have and what disqualifies them; see the matching companies live. A prospect list is a query you can explain, not a file.
  • Native mail providers. Gmail (REST + OAuth) and any SMTP+IMAP mailbox.
  • The outbox approval ladder. Every outgoing email waits for approval by default (pending_review → approved → sent), reviewed one at a time in a decisions desk with A/R/J/K. A deterministic pre-send gate — no model involved — rejects unrendered merge tags, banned phrases, over-length bodies, too many links, HTML, undeliverable addresses and recipient mismatches.
  • Bounce handling and a kill switch. A bounce marks the address invalid, cancels that prospect's queued sends, and trips a global pause past a configured bounce rate. harvey sending pause|resume.
  • Claude usage tracking. Per-call attribution by agent, task and model, plus live subscription-quota gauges read the same way /usage does, so Harvey throttles itself and leaves headroom for your own interactive work. harvey usage.
  • CSV export. harvey export writes a sequencer-ready list. Harvey is worth running as a list-builder even if you never let it send.
  • A dashboard worth looking at. Opens on Today — what needs a human — with an activity feed and a rail of pipeline figures, collector runs and setup. Light and dark are the same design re-tokenised; appearance follows your OS by default. Fonts are vendored so it renders with the network off.
  • A Python version guard. python3 -m venv on macOS builds from the system 3.9, which cannot run Harvey. It now fails immediately with the fix instead of surfacing later as an unrelated-looking ImportError.

Changed

  • The data model. Every fact Harvey learns is now an observation — a row in observations, never a column. A new signal needs no migration, re-observing over time is a free time series, and confidence and provenance travel with the fact. The vocabulary is governed: a database trigger rejects any signal code not in signal_codes.
  • Email finding is pattern-first. Harvey derives the domain's address format and verifies one candidate rather than brute-forcing name variations, and grades every address honestly as verified / risky / guess / invalid. Only deliverable addresses are ever sent to.
  • The default mail provider is now gmail. Instantly is still supported as a legacy option.
  • Discovery depth defaults to 30, not 100. Google removed 100-results-per-page in September 2025, so depth 100 is now billed as ten pages nearly everywhere — and below rank 30 it is mostly directories.
  • The dashboard's HTML, CSS and JS are real files in harvey/web/ instead of a 1,500-line string literal in dashboard.py. Still no build step.

Fixed

  • Guessed emails were being reported as verified whenever the domain merely had an MX record. The migration downgrades every legacy email_verified=1 row to guess.
  • Cohorts counted the wrong companies. PROFILE records a false boolean when it checks and finds nothing — which is correct — but cohort() matched on the row existing, so "companies running Google Ads" silently meant "companies we checked for Google Ads", which was all of them. Cohorts now read the newest observation per company and signal, and require it to be true.
  • Usage tracking counted every Claude session on the machine, not just Harvey's. The transcript reconciler is gone; the ledger contains only rows Harvey's own Brain wrote.
  • Page titles were being extracted as people ("Request Service", "Commercial Services"), and titles bled between team members.
  • Confirming a signal before anything had loaded the catalog reported success while changing nothing.
  • The free OSM provider answered unmapped search terms with every registered office in the area — coffee roasters and churches for a roofing query. It now refuses the query and names the trades it can search.

Security

  • Test fixtures no longer carry real businesses' phone numbers or street addresses.
  • harvey.local.yaml (gitignored) overrides the tracked harvey.yaml, so a fork can hold real product configuration without it reaching a public repo.

[0.1.0]

Initial release: the heartbeat loop, five sub-agents, the skills library, the website trainer, Instantly-based sending, and the first dashboard.