agentgov for IT, from the first sign-in to the audit

Start

agentgov gives IT one place for the rules every employee's AI agent follows, and one page that shows who is connected and what their agents report. It is free. You need a work email; your people need the AI app they already use. Open the dashboard and sign in with the six-digit code we mail you.

The limit, plainly. Rules plus watching, not a lock. agentgov tells every agent the company rules and shows IT what happened. Where IT deploys our settings files, the app itself enforces some rules. A person can still open another chat, another app or another computer.

1 Create your company

After the first sign-in, give your company a name. You become its first admin. We mail you the three next steps: switch on rules, invite people, and, if you like, deploy the settings files. You can be in more than one company; each has its own rules, people and activity, stored apart.

2 Switch on rules

On Rules, pick a bundle: Recommended start, Strict, with customer data or Developers. It ticks its rules; nothing is saved until you press Save rules. Then add or drop single rules, fill in the lists (allowed AI tools and connectors, data classes, models, websites, what needs your approval) and write up to 30 rules of your own. Every save is a new version with what changed; older versions stay readable.

Each rule shows how far it goes in each app: enforced by the app, partly enforced (with what still gets through), reported by the app, set in the app's admin console, or told to the agent and reported. All 22 are on the rule sets page.

3 Invite your people

On People, paste up to 50 addresses and press Send invites. Each person gets one mail with a link that works once, for 14 days. The mail and the join page say plainly what you will see and what you never see. When an address is not on your own domain, the mail says so, so nobody joins the wrong company by mistake.

The person opens the link, presses Join, picks their app and follows a few steps. The page turns green by itself once their app has signed in. If they get stuck, they press Tell IT I'm stuck and you see it under "Needs a hand". Resend or revoke an invitation from the same page.

4 Set up computers (optional)

On company computers with Claude Code, Codex or Cursor you can go further. On Connect → Your computers:

1. Make a reporting key. Give it a label such as "Laptops". It is shown once. It can only write activity to your company: it cannot read anything, see people or rules, or touch another company.
2. Download the reporting script with the key filled in: agentgov-report.sh (macOS, Linux) or agentgov-report.ps1 (Windows). Deploy it with your device management. It reads the app's report on the computer, keeps only the names (the event, the tool, the connector, the session, the person) and sends nothing else.
3. Tell the script who uses the computer. Set AGENTGOV_USER to the person's work email on each computer, or download the script with {{EMAIL}} in it for device management that fills in files per user.
4. Download the settings files for each app and deploy them to the paths on Connect. They are written from your rules and hold no key. When you change rules you want enforced, download and deploy them again.
5. Check it: run the script with --test on one computer; it says whether the key works. Then start the app there: the person, once joined, shows "Reported by a computer" on Live.

Activity from a computer is kept only for people who joined. Rows from computers carry the key's label, and rule hits from computers are marked "not confirmed", because anyone who copies a key could send them. Retire a key at once, or with a week's overlap while you deploy a new one.

5 The dashboard

Overview shows who is connected now, who is not, requests waiting and rule hits today, and what needs a hand. People lists everyone with their status and apps. Live shows who is doing what right now. Activity has every event with filters and a download; its second tab, the admin log, lists what admins did. Rules, Requests, Connect and Settings do what they say. A tour with pictures is on the dashboard page. Every Monday, admins get a digest: who is not connected, who was quiet, rule hits, requests.

6 Requests and approvals

When a rule says "ask IT first", or the work falls in a category you marked (research, a new feature, sharing outside, automation, customer data), the agent sends you a request with what, why, which tools and which data, and tells the person it is waiting. You get a mail. On Requests you approve, approve with conditions, or deny, with a note. The agent follows your answer and your conditions. When the request is an idea, the agent can also hand it to dataready to test it on a synthetic community before anyone builds it.

7 What the agent is told

At the start of each session the agent reads your rules as short lines, in this order: the rules you switched on, your own rules, then your lists.

MUST    Keep the agentgov connector on and read the company rules…
NEVER   Never open .env files, private keys or credential stores…
ASK IT  Before building a bot, an agent, a scheduled job or an integration…
REPORT  Call team_activity_report when work starts…

It reports when work starts, around steps a rule marks "ask", and when work ends, with one plain sentence about the kind of work, at most 140 characters: never names, addresses, numbers, links, passwords or the person's own words. After reporting it tells the person in one line what IT will see. If a person asks it not to report, it tells them it is the company's rule, and then does what they decide.

One rule, "Mark AI-made customer content", asks the agent to say when text that reaches customers was written with AI help; the EU AI Act asks for this in some cases. A code reading, not a legal opinion.

8 What IT sees, and what it never sees

The same two lists are on every invitation and join page.

App by app:

9 How long data is kept, and deleting

• Activity: 180 days at most. Choose 90 or 30 in Settings.
• Requests, decisions and the admin log: 3 years.
• "I'm stuck" notes: 90 days.
• Invitations: until accepted; the address is blanked 30 days after an invitation expired or was revoked.
• The organisation: while it exists. "Delete the organisation" in Settings deletes everything at once.

Everything is stored in the EU. Each person sees what you see about them on their own page (their latest activity, requests and notes) and can download all of it there. When someone leaves, their agent stops reading your rules and stops reporting; activity from their computer is no longer kept under their name.

10 Limits, per app

• Claude Code (tested): with your settings file, rules marked "enforced" are enforced by the app; those marked "partly" stop the common forms of a command, and each says what still gets through.
• Codex: requirements and hooks come from Codex's documentation and are not yet checked on a real account. Hosted tools such as web search are not covered by its hooks.
• Cursor: our hooks only watch, after a shell run or a connector call; they never block. Admin-console settings on Cursor's business plans are yours to set.
• Claude and ChatGPT: rules are told to the agent and it reports what it does. On team and business plans an owner or admin may have to allow the connector first. ChatGPT may ask the person to confirm each report; then the agent reports only the start and the end.
• VS Code, Windsurf and other apps: rules told, work reported, nothing enforced by the app.

For every app but Claude Code: steps not yet checked on every plan.

For employees

Your IT team uses agentgov to share the company's rules for AI tools. You connect it once to the AI app you already use; then your agent reads the rules at the start of each session, asks IT before work that needs approval, and tells IT in one short sentence what kind of work it is doing. Never your messages, files or what you paste.

From your own page you can see what IT sees about you (your latest activity, requests and notes), download all of it, switch between companies, or leave. If something does not work, press "Tell IT I'm stuck" on your join page; don't paste links, codes or error text with your details, IT will ask.

FAQ