RevenueOS runs free as open-source revenue agent, charging only for approved measured results

An open-source project called RevenueOS landed on Hacker News's front page on 13 September 2026, pitching itself as an automated revenue department for a single business. The software, self-hosted and MIT-licensed, connects to a company's website and accounts, surfaces revenue-related actions, and executes them only after a human approves — with payment requested only after a measured result the user agrees with.

How the agent works

According to the project's README on GitHub, RevenueOS operates in five stages. After an initial connection, it infers what it can about the business from the website and labels every guess it makes. It then reads the site, advertising accounts, leads and mailbox to produce a short list of candidate actions. Nothing changes until the owner approves an action. Approved work can include deploying a site fix, pausing a campaign that is wasting budget, sending outreach email, booking calls or raising invoices. Finally, the system re-checks the world and records a before-and-after result next to each action.

The payment model

The pricing claim is the project's most distinctive element. Every worker, the control panel and the executors run free of charge. Once RevenueOS has measured a result on an action the user approved, it keeps running without charge for a further 14 days; after that, continuous operation requires a Pro tier at $99 per month. One-shot runs and the panel never lock, and there is no trial clock that starts before a result has been seen, the README states.

Setup and technical requirements

Installation is a single pip install, and a demo command can run against any URL. The demo builds its workspace in a temporary directory and deletes it on exit, and a --no-llm flag keeps it entirely local. When run against the neutral example.com domain, the demo reported four findings: a missing meta description, thin page content, no sitemap and no canonical URL. A GitHub Action variant can add the check to any workflow.

The project requires Python 3.12+ and Node 20+. An LLM is optional — an Anthropic API key or a signed-in Claude Code CLI — and without one, the README says the deterministic half still runs: the site crawl, ad-export waste and pacing checks, search-term waste, lead qualification, inbox handling, templated outreach drafts and all measurement. Skill-based deliverable execution, the ads control audit, the relevance gate and specialist roles explicitly need a model. Site deploys, campaign pauses, invoices, bookings and email sends do not.

Integrations and workers

Connections cover Stripe, Google services (Search Console, GA4, Calendar, Ads), Meta Ads, git-hosted sites and WordPress. All start read-only until the user enables changes, and every change still waits for approval in the panel, which is served locally on port 8791 with password-protected sessions. Eleven scheduled workers handle SEO, ads auditing (including a pass/fail evaluation of 414 controls across 12 platforms), live ad account monitoring, analytics, billing, prospect discovery, outreach, inbox triage, content publishing, external monitoring and measurement.

A documented proof run

The README describes a verified run on 12 September 2026 against Plausible Analytics, connected read-only using only its public web presence. The system crawled 12 pages, checked competitor authority, searched Hacker News through a relevance gate and queued six content opportunities. One was approved; Claude executed an SEO content-brief skill in 1 minute 23 seconds and produced a 14 KB deliverable, and the outcome deliverable_written moved from 0 to 1 in the results view. The full transcript is published in the repository's documentation. The project states it reports only measured outcomes and does not claim revenue it has not measured.

Why it matters

Most agent products charge by seat, token or subscription regardless of value delivered. RevenueOS ties payment to a measured, user-approved outcome, an approach that could pressure pricing norms in agentic software if it scales. The approval model also addresses a real trust problem: agents that send email or touch ad budgets are risky to leave unattended, and read-only-by-default connections plus per-action sign-off give operators a brake. Being open source and self-hosted makes the claims auditable — though in practice the pay-on-result model becomes a $99 monthly subscription for continuous use once a first result is agreed.