agentgovConnect
Connect agentgov to the app you already use
Your people do it once, in a few minutes, from their invitation. IT can go further on company computers with the settings files agentgov writes.
For your people
Each person opens the link in their invitation; the join page shows the steps for their app and turns green once it is connected. The same steps are here. Sign in with your work email, the one IT invited.
https://agentgov-mcp.adnan-c3a.workers.dev/mcp- Open Claude (the app or claude.ai) and go to Settings, then Connectors.
- Click Add custom connector.
- Type the name
agentgovand paste the address below. - Click Add, then Connect.
- Sign in with your work email: type it, then open the link we send you (or type the 6-digit code).
https://agentgov-mcp.adnan-c3a.workers.dev/mcpOn a Team or Enterprise plan an owner may have to add it for everyone first. No Add custom connector in your settings? Ask IT.
Steps not yet checked on every plan. If a menu is named differently, tell IT from your invitation page.
- Open chatgpt.com and go to Settings, then Apps (or Apps & Connectors), then Advanced settings. Turn on Developer mode.
- Click Create. Type the name
agentgovand paste the address below. Keep the suggested sign-in choice and tick “I trust this application”. - Click Create and sign in with your work email when it asks.
- In a new chat, click + and pick
agentgov.
https://agentgov-mcp.adnan-c3a.workers.dev/mcpThis needs a paid ChatGPT plan, on the website. On Business and Enterprise plans an admin may have to allow it first. ChatGPT may ask you to confirm each time your agent tells IT what it is working on; that is expected.
Steps not yet checked on every plan. If a menu is named differently, tell IT from your invitation page.
- Open your terminal and run the line below.
- Start Claude Code by typing
claude. - Type
/mcp, pickagentgovand choose Authenticate. - Sign in with your work email in the browser tab that opens.
claude mcp add --transport http agentgov https://agentgov-mcp.adnan-c3a.workers.dev/mcp --scope user- Open your terminal and run the two lines below, one after the other.
- Sign in with your work email in the browser tab the second line opens.
codex mcp add agentgov --url https://agentgov-mcp.adnan-c3a.workers.dev/mcp
codex mcp login agentgovCodex does not know mcp add? Update Codex, or add the two lines below to ~/.codex/config.toml (on Windows %USERPROFILE%\.codex\config.toml; create the file if it is not there), save, and run the second line above.
[mcp_servers."agentgov"]
url = "https://agentgov-mcp.adnan-c3a.workers.dev/mcp"The sign-in is refused? Update Codex to its newest version and run the second line again.
Steps not yet checked on every plan. If a menu is named differently, tell IT from your invitation page.
- Click Add to Cursor. Cursor opens.
- Click Install.
- Click Connect next to
agentgovand sign in with your work email.
https://agentgov-mcp.adnan-c3a.workers.dev/mcpAdd to Cursor Or open the install page
The button does nothing? Cursor isn't installed on this computer. In Cursor's settings you can also add it by hand with the address below.
Steps not yet checked on every plan. If a menu is named differently, tell IT from your invitation page.
- Click Add to VS Code. VS Code opens.
- Click Install.
- When it asks, click Allow and sign in with your work email.
- Use
agentgovfrom the Copilot chat in Agent mode.
https://agentgov-mcp.adnan-c3a.workers.dev/mcpAdd to VS Code Or open the install page
The button does nothing? VS Code isn't installed on this computer.
Steps not yet checked on every plan. If a menu is named differently, tell IT from your invitation page.
- In Windsurf, open Settings, then Cascade, then the raw settings file (
~/.codeium/windsurf/mcp_config.json). - Paste the lines below and save. If the file already lists other servers, add
agentgovnext to them. - Refresh the list and sign in with your work email when it asks.
{
"mcpServers": {
"agentgov": {
"serverUrl": "https://agentgov-mcp.adnan-c3a.workers.dev/mcp"
}
}
}Steps not yet checked on every plan. If a menu is named differently, tell IT from your invitation page.
- In your AI app's settings, look for Connectors, Tools or Integrations.
- Choose to add a custom one, name it
agentgovand paste the address below. - Sign in with your work email when it asks.
https://agentgov-mcp.adnan-c3a.workers.dev/mcpNot sure which app you use, or nothing fits? IT can help.
Steps not yet checked on every plan. If a menu is named differently, tell IT from your invitation page.
Then say company rules to your agent. For work that needs approval, say I have an idea for … and it asks IT for you.
For IT: your computers
On company computers with Claude Code, Codex or Cursor, IT can deploy settings files the way it deploys other settings. Then the app itself enforces some rules, and reports each session and tool by name, even when the agent forgets to.
1. A reporting key
Make one on the dashboard's Connect page; it is shown once. It can only write activity to your organisation: it cannot read anything, see people or rules, or touch another organisation. Someone who copies it from a computer could send activity under a false name, so retire it at once or with a week's overlap when a computer leaves.
2. The reporting script
agentgov-report.sh for macOS and Linux, agentgov-report.ps1 for Windows. The app runs it at each session start and before each tool. It reads the app's report on the computer, keeps only the names below and sends them in the background (on Windows it waits at most two seconds); it never stops the app.
3. The settings files
Generated from your rules: Claude Code's managed-settings.json, Codex's requirements.toml and hooks, Cursor's hooks.json. They hold no key. Download them from the dashboard after each rules change you want enforced.
What the script sends
- the event: a session started or ended, a tool is about to run
- the tool's name and its connector's name
- the session's id and the permission mode
- which app, and the person's email as the computer knows it (set per computer by IT)
What it never sends
- prompts, answers or the chat
- file contents or file names
- commands and their details, tool inputs and outputs
- anything else: our server refuses any other field, and keeps nothing for a person who has not joined
Where the files go
| File | macOS | Linux | Windows |
|---|---|---|---|
| The reporting script | /Library/Application Support/agentgov/agentgov-report.sh | /opt/agentgov/agentgov-report.sh | C:\ProgramData\agentgov\agentgov-report.ps1 |
Claude Code managed-settings.json | /Library/Application Support/ClaudeCode/ | /etc/claude-code/ | C:\Program Files\ClaudeCode\ |
Codex requirements.toml | your device management, or /etc/codex/ | /etc/codex/ | %ProgramData%\OpenAI\Codex\ |
Cursor hooks.json | /Library/Application Support/Cursor/ | /etc/cursor/ | C:\ProgramData\Cursor\ |
Claude Code
Tested. The settings file adds agentgov for everyone, enforces the rules marked "enforced" and partly enforces those marked "partly", and reports every session and tool by name.
Codex
Requirements and hooks from Codex's documentation.
Not yet checked on a real account: the hook names and where managed hooks go.Cursor
Hooks that only watch: session start and end, shell runs and connector calls after they ran, by name. They never block anything.
Not yet checked on a real account: the connector's name in each call.Steps not yet checked on every plan for Codex and Cursor. Each person's email comes from the computer, set by IT, so a computer's report is shown with its key's label and a rule hit from a computer is marked "not confirmed".
What each app lets agentgov see
How much IT sees depends on the app, and on whether IT deployed the settings files.
| App | agentgov sees | agentgov never sees | Rules enforced by the app |
|---|---|---|---|
| Claude Code + IT's managed settings | every session start and end; the name of every tool the agent uses (built-in and connectors) and its server; the permission mode; calls to agentgov; the agent's short reports | prompts, answers, file contents, commands, tool inputs and outputs — the reporting script reads the app's report on the computer, keeps only the names and sends nothing else | the rules marked Enforced (allowed connectors, models, the skip-all-permissions mode, company-only settings); those marked Partly enforced in part (common forms of a command are stopped, other forms are not) |
| Codex + IT's requirements and hooks (not yet checked on a real account) | session start and end; tool names; calls to agentgov; the agent's short reports | the same as above; hosted tools such as web search are not covered by Codex hooks | the rules marked Enforced or Partly enforced for Codex (approval policies, sandbox modes, the allowed connectors) |
| Cursor + IT's hooks file | session start and end; shell runs and connector calls after they ran (names only); calls to agentgov; the agent's short reports | the same as above | none generated (agentgov's Cursor hooks only watch); settings in Cursor's admin console on Enterprise plans |
| claude.ai, Claude desktop, ChatGPT | the connection, every call to agentgov, the agent's short reports | everything else the agent does: other connectors, files, web search, the chat | only what the app's own admin console offers; agentgov shows the steps |
| VS Code, Windsurf and other apps | the connection, every call to agentgov, the agent's short reports | everything else | none |
| Any app without agentgov connected | nothing; the person shows as not connected | everything | none |
The limit, plainly
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.