agentgovFree · runs on your computer · in Claude Code, Cursor or Codex
Your app has AI agents. Do you know what they may do?
agentgov reads your code and writes down every model call, agent, MCP server and tool, the ones you built and the ones someone else built that you plug in. It shows where one of them can delete, pay or send with nobody asking, where a stranger's text flows into its instructions, and what the AI Act will want from you. With a score, a file and a line. Then it puts them under control.
Example: a registry for a shop app with a score of 62, grade C. A refund tool from your own code needs approval, a delete tool from someone else's MCP server is forbidden, a search tool is allowed, and a send-email tool has no policy yet, so the release gate holds the release.
22 seconds
See it work.
The check reads the code, finds what an agent may do, and GOV on puts a policy on every tool. A code reading, not a legal opinion.
Yours, and everyone else's
Not only the agents you build. The ones you plug in, too.
Most agents at work were built by someone else: an MCP server from the internet, a vendor's tool, the coding agent itself in Codex, Claude Code or Cursor. agentgov covers those the same way.
Every server you plug in
It reads your project's MCP settings and lists each server you added, with the tools it brings and what they may do: read, write, pay, send, delete.
A policy for someone else's tool
Allow it, ask first or forbid it, with the reason, in the same registry as your own tools. The release gate counts them too.
Rules inside Codex and the rest
With Teams, IT sets the rules once and each person's agent follows them in Codex, Claude Code or Cursor, asking before the work starts. Rules plus watching, not a lock.
From a check
What it finds, in plain words
An app with a support chatbot that can refund a customer. agentgov read the code and said this.
Your agent can refund any amount with nobody confirming it
The tool refund_customer takes a customer id and an amount, moves money and is wired to the chat agent with no confirmation step and no cap. The customer's own message goes straight into the agent's instructions.
Nobody can tell afterwards who asked for what
The model calls and the tool calls are not logged with the person behind them. After a bad refund there is no trail to read.
src/agents/support.ts:12 What it costs: an incident you cannot explain, to a customer or to an auditor.Every finding has its consequence, its remedy and the exact file and line. A code reading, not a legal opinion.
How it works
Four steps, in plain words
You talk to your coding agent the way you always do. agentgov does the reading and the counting on your computer; your agent does the judging with the model you already use.
-
Free
The inventory
It reads your app's code and writes down what is there: every model call, agent, MCP server, tool and prompt, what each one can do, and where. Your agent fills in what needs judgement.
check my agents -
Free
The checks
Tools that delete, pay or send with no person confirming. Untrusted text that flows into a prompt. Secrets in prompts. No log of who did what. No limit on steps or spend. The AI Act tier. A score 0–100 and a grade.
run the checks -
Free
Live
It starts your MCP server or agent from the code on your computer, after your yes, connects as a client and calls the tools for real: does a guarded tool run without confirmation, does a planted instruction get followed, does a result leak a secret?
test my agents live -
GOV on
Under control
A registry of every agent and tool with a policy: allowed, needs a person's approval, forbidden. A release gate that stops a new dangerous tool with no policy. A decision register. Documentation for an audit. Fix, check again, repeat.
fix it
Nothing but counts, rule names, titles and file:line references ever leaves your computer, and only when you say share the report. Your code, your prompts and your keys stay where they are.
Two ways in
The check is free. GOV on is free too.
Everything you need to know where you stand.
- The inventory of every agent, MCP server, tool and prompt
- The checks: a score, a grade, every finding with file and line
- The live test of your MCP server or agent, on your computer
- The registry view, the AI Act tier, the decisions you made
- Teams: your company's rules, requests and the inbox
- The report: one link and a code, to send to a client or a boss
The same tool fixes it, with a trail.
- A policy per tool: allowed, needs approval, forbidden, with the reason
- The release gate: a dangerous tool with no policy does not ship
- The fix plan, in order, and the re-check until the score is good
- The decision register: who decided what, when and why
- Documentation for an audit and the AI Act, generated from all of it
- Unlimited projects, no card
Teams · free
Your IT department sets the rules. Every employee's agent follows them.
Matej in IT writes the rules once: which AI tools and MCP servers are allowed, which data may go into an AI, what needs his approval. Nina in marketing gets one mail with the install line. When she says "I have an idea", her agent asks Matej first, waits for his yes, follows his conditions, and hands the idea to dataready to see whether it is worth building.
Illustration: Nina at her laptop has an idea; a request travels as an envelope past the rules sheet into Matej's inbox, and he answers ok.
Rules, once
Allowed tools and servers, data classes, what needs approval. One page, and they reach everyone.
Ask before, not after
Research, a new feature, sharing outside, automation, customer data: the agent asks first and waits.
Rules plus watching
A watcher tells IT which unapproved MCP servers appear on people's computers. Names and versions, never contents. Not a lock: a person can still open another chat.
Who it is for
You put an agent in your app. Now someone will ask what it may do.
Founders and solo builders
Built with Claude Code, Cursor, Lovable or Replit, and now there is a chatbot with tools in front of paying customers. Know what it can do before a customer finds out.
Dev shops
You ship apps with agents to clients. Send them the report with every finding and its file, and the registry that says what each tool may do.
IT and compliance
Someone has asked about the AI Act, a security review or an audit. The registry, the decision register and the documentation come out of the same tool.
Install · one line
Add it to the tool you already use
A small MCP server your coding agent starts by itself, inside your app's folder. Pick your tool, copy the line, sign in with your email once. The check is free.
npx -y --package=https://agentgov.si/dl/agentgov-launcher.tgz agentgov-mcp
claude mcp add agentgov -- npx -y --package=https://agentgov.si/dl/agentgov-launcher.tgz agentgov-mcpFor your whole team (writes .mcp.json): claude mcp add -s project agentgov -- npx -y --package=https://agentgov.si/dl/agentgov-launcher.tgz agentgov-mcp. Then say check my agents; the first time, it opens a sign-in page.
{"mcpServers":{"agentgov":{"command":"npx","args":["-y","--package=https://agentgov.si/dl/agentgov-launcher.tgz","agentgov-mcp","--root","/path/to/your/app"]}}}Settings → Developer → Edit config. Put your app's folder after --root, then restart Claude Desktop.
{"mcpServers":{"agentgov":{"command":"npx","args":["-y","--package=https://agentgov.si/dl/agentgov-launcher.tgz","agentgov-mcp"]}}}Save the file in your project; Cursor tells agentgov which folder it is working in.
codex mcp add agentgov -- npx -y --package=https://agentgov.si/dl/agentgov-launcher.tgz agentgov-mcpOr in ~/.codex/config.toml: [mcp_servers.agentgov], command = "npx", args = ["-y", "--package=https://agentgov.si/dl/agentgov-launcher.tgz", "agentgov-mcp", "--root", "/path/to/your/app"].
{"mcpServers":{"agentgov":{"command":"npx","args":["-y","--package=https://agentgov.si/dl/agentgov-launcher.tgz","agentgov-mcp","--root","/path/to/your/app"]}}}Settings → Cascade → MCP servers → View raw config, paste, put your app's folder after --root, refresh.
{"mcpServers":{"agentgov":{"command":"npx","args":["-y","--package=https://agentgov.si/dl/agentgov-launcher.tgz","agentgov-mcp","--root","/path/to/your/app"]}}}MCP Servers → Configure → paste, with your app's folder after --root.
Needs Node.js 22.13 or newer. Install once: agentgov updates itself when your editor starts it. On Windows, use "command": "cmd" with "args": ["/c", "npx", "-y", "--package=https://agentgov.si/dl/agentgov-launcher.tgz", "agentgov-mcp", …]. Teams on claude.ai or ChatGPT, with nothing to install: the connector https://mcp.agentgov.si/mcp (how).
Built with prodready? Tested with dataready? Govern it with agentgov.
prodready makes what you built safe to put in front of real users. dataready tells you which of those users will stay. agentgov tells you what your AI agents may do, and keeps it that way. Same kind of install, made by the same team.
Questions
Good questions, short answers
What is agentgov, in one sentence?
A tool in your coding agent that reads your app's code, tells you what its AI agents may do and where that is dangerous, and then puts them under control: policies, a release gate, a decision register and documentation.
Is the check really free?
Yes. The inventory, the checks, the live test, the registry view, Teams and the report cost nothing, as often as you like. GOV on, the part that fixes things with a trail, is free too: sign in by email and it is on. No card. See pricing.
Do you see my code?
No. agentgov runs on your computer, reads only your app's folder (never your .env files or keys) and keeps what it makes in .agentgov/ inside your project. From us it downloads the know-how: what to look for and how to grade it. The one exception is yours to make: share the report shows you exactly what would go (counts, titles, file names with line numbers, never code or file contents) and uploads it only after you agree.
Does it use AI?
Your coding agent does, with the model you already use. agentgov itself does the reading and the counting; when a finding needs judgement (does a person really confirm this refund?), your own agent reads the code and answers. We run no model of our own.
Is it legal advice about the AI Act?
No. agentgov reads the code, says which risk tier the system probably falls in and which obligations and documents you are missing, and writes the first drafts. A code reading, not a legal opinion: a first opinion to take to a lawyer. Details on the AI Act page.
Will the live test touch my real app?
No. It starts your server or agent from the code on your computer, in a test-mode environment, after you say yes once; a private test address also works, production addresses are refused. Tools that delete, pay or send are called only when you list them.
Which tools does it work with?
Claude Code, Claude Desktop, Cursor, Codex, Windsurf and Cline, on macOS, Linux and Windows, with Node.js 22.13 or newer. Teams also works from claude.ai and ChatGPT through a connector, with nothing to install.
Can Teams stop an employee from using some AI?
Not by force. agentgov makes the employee's agent follow the rules and ask before it starts, and the watcher tells IT which unapproved MCP servers appear on people's computers. A person can still open another chat window. That is rules plus watching, not a lock; a lock is your IT department's own systems.
What do you keep?
Your sign-in email, the reports you choose to share (until you delete them), your organisation's rules and requests while it exists, and anonymous page counts. Details are on the privacy page.
A code reading, not a legal opinion.