Six agents that do real work and answer to you.
Each one is a Claude Code agent you run on your own computer, built for a specific job and boxed in by a policy you can read. They work from the exports you already have, write their results to a folder, and leave every change to a person.
Join the Academy to get all sixWhat makes them governed
The six agents
AI Spend Auditor
- Reads your AI bills, card exports and seat lists
- Finds idle seats, duplicate tools and personal plans on company cards
- Ranks cut, consolidate and keep, with the monthly saving and the evidence row for each
- Writes a one-page summary for the CFO and a cancellation checklist for a person to work through
Never: Cancels a subscription, emails a vendor or logs in anywhere.
Agent Authority Reviewer
- Checks a repository's agent setup against 74 rules
- Lists every place an agent can act with no identity check, approval or independent CI gate
- Gives each finding a file and line, plus the fix as a patch
- Builds an evidence package with checksums for an auditor
Never: Applies a patch, pushes code or calls a cloud API.
Month-End Close Assistant
- Categorizes every transaction by rules you approve
- Matches Stripe, Shopify and PayPal payouts and invoice payments to deposits
- Flags duplicates, new vendors and unusual amounts
- Builds a cash-basis P&L that ties out, plus import-ready CSVs for QuickBooks or Xero
Never: Logs in to your bank or posts to your accounting software.
Ad Account Analyst
- Reads Ads Manager and Google Ads CSV exports
- Writes a daily brief against the trailing week
- Finds creative fatigue and ranks what to kill, scale and test by money at stake
- Runs a weekly audit scorecard and builds a testing plan
Never: Touches your ad accounts or changes a budget, bid or status.
Support Reply Drafter
- Reads Zendesk, Gorgias, Intercom, Help Scout or Gmail exports
- Masks customer names, emails and card numbers before the model sees them
- Drafts replies in your tone, citing the policy passage behind each one
- Flags legal threats, refunds over your limit and anything your policies don't cover
Never: Sends a reply or changes a ticket.
Competitor & Market Brief
- Reads competitors' public pages from an allowlist you control, respecting robots.txt
- Links every claim to its source page, the date it was read and a confidence label
- Tracks what changed since the last brief: prices, pages, open roles
- Blocks publishing until every claim passes its citation check
Never: Logs in, submits forms or reads a site you haven't allowed.
What you need
- Claude Code on a Mac, Windows or Linux computer, with your own Claude plan.
- Python 3.9 or newer. Each agent's INSTALL.md walks through setup step by step.
- The exports for the job, such as bank CSVs or an ad account export. Each agent lists exactly which.
Members download all six from their purchase page. Each zip includes the agent, its tools and tests, sample data, worked examples and a SECURITY.md that spells out what it can and can't do. The commercial license comes with the membership; the terms are in the LICENSE file inside each zip.
Get all six with the Academy
Membership includes the 30 video lessons on making agents safe to run at work, plus these six agents.
See the Academy