MCP server

One line. It runs on your computer.

agentgov is a local MCP server: your coding agent starts it inside your app's folder. It reads the code there, keeps everything it makes in .agentgov/, and fetches from us only the know-how: what to look for and how to grade it. Your coding agent does the judging.

Install

Package npx -y --package=https://agentgov.si/dl/agentgov-launcher.tgz agentgov-mcp

Needs Node.js 22.13 or newer. The first time you use it, it opens a sign-in page in your browser. Type your email, click the link we send (or type the six-digit code from the same mail), and confirm. No password, no card. The sign-in is kept in your user settings folder; your app's data stays in .agentgov/ in your project.

Claude Codein your app's folder

claude mcp add agentgov -- npx -y --package=https://agentgov.si/dl/agentgov-launcher.tgz agentgov-mcp

For 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.

Claude Desktopclaude_desktop_config.json

{"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.

Cursor.cursor/mcp.json in your app

{"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.

Codexin your app's folder

codex mcp add agentgov -- npx -y --package=https://agentgov.si/dl/agentgov-launcher.tgz agentgov-mcp

Or 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"].

Windsurfmcp_config.json

{"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.

Clinecline_mcp_settings.json

{"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.

On Windows, use "command": "cmd" with "args": ["/c", "npx", "-y", "--package=https://agentgov.si/dl/agentgov-launcher.tgz", "agentgov-mcp", …]. Needs Node.js 22.13 or newer. You install it once: agentgov checks for a new signed release each time your editor starts it, so restarting the editor is how you update. The launcher package's sha256 is 756e7d41d2df4a87e06716dddad98e6c55a0316d150e313e077bcf50863b165d; each engine release is signed and checked before it runs.

Teams on claude.ai or ChatGPT: a connector, nothing to install

The rules your IT department set, your requests and the admin's inbox also live on our server, so an employee who works in claude.ai, Claude Desktop or ChatGPT needs no local install: add the address as a custom connector (Claude: Settings → Connectors → Add custom connector), sign in with your email, and say company rules or I have an idea. The inventory and the checks of an app stay in the local app above; this connector carries only Teams.

Teams connectorconnector URL

https://mcp.agentgov.si/mcp

Claude Code, as a second server next to the local one: claude mcp add --transport http agentgov-teams https://mcp.agentgov.si/mcp --scope user.

Just ask

Things you can say

No commands to learn. Say it in your own words, in any language; your agent picks the right tool.

You say	What happens
`check my agents`	agentgov reads your app on your computer and drafts the Agent Inventory: every model call, agent, MCP server, tool, prompt and data flow, each with its file and line. Your agent completes what needs judgement.
`confirm the inventory`	Fix what it lists as unknown, then lock the inventory in as a version.
`run the checks`	The danger spots: a score 0–100, a grade A–F, every finding with its consequence, remedy and file:line. What needs reading becomes a short task for your agent.
`start my app`	It starts your server or agent from the code on 127.0.0.1, in a test-mode environment: it shows the exact commands first and runs them only after your yes. Never production.
`test my agents live`	It connects to the running server as a client and calls the tools for real: does a guarded tool run without a confirmation, does a planted instruction in a result get followed, does a result leak a secret? Tools that delete, pay or send are called only when you list them.
`show the registry`	Every agent and tool with its policy: allowed, needs approval, forbidden, or not decided yet.
`forbid delete_user` · `refund needs approval`	A policy per tool, with the reason, recorded as a decision. GOV on.
`run the release gate`	Pass or blocked, with the reasons: a dangerous tool with no policy, an open blocker, a forbidden tool still wired, an inventory newer than the registry. GOV on.
`fix it`	The fix plan, in order, with the remedy and the files for each finding. Your agent makes the changes; agentgov never edits your project. Then check again. GOV on.
`I fixed the refund finding` · `accept the risk`	Marks a finding fixed with your evidence (the next check verifies it) or accepted with a reason. Both become decisions. GOV on.
`write the audit documentation`	The registry, the risk assessment, the AI Act document and the change log, as markdown in `.agentgov/docs/`, yours to commit. GOV on.
`company rules`	The rules your IT department set for AI tools, MCP servers and data, and what needs approval.
`I have an idea for …`	Your agent sends the request to your organisation's admins and waits. When approved, it follows the conditions and hands the idea to dataready to see whether it is worth building.
`inbox`	For the admin: the requests waiting; approve, deny or approve with conditions. The member is mailed.
`share the report`	One link and an access code on agentgov.si for this project's report. It shows you exactly what would go and uploads nothing until you agree.
`open the dashboard`	A private page on your computer with the inventory, the findings, the registry, the decisions and the gates.
`status`	Signed in or not, the licence, your project folder, your local data and the know-how versions.
`help`	What agentgov can do, and where to start.

Prefer slash commands? /agentgov:scan, /agentgov:check, /agentgov:live, /agentgov:run, /agentgov:registry, /agentgov:fix, /agentgov:gate, /agentgov:docs, /agentgov:team, /agentgov:request, /agentgov:inbox, /agentgov:status, /agentgov:dashboard, /agentgov:report and /agentgov:help do the same.

Under the hood

The tools

Thirty-six tools, all on your computer. Your agent picks them; you never have to name one. The ones marked GOV on need the licence; everything else is free.

agent_scanfree

Reads your app's folder (read only, never .env files or keys) and drafts the Agent Inventory; hands your agent one short task for what needs judgement.

agent_inventory_setfree

Saves corrections and confirms the inventory as a new version, with what changed.

gov_checkfree

Runs the checks over the inventory: a score, a grade, every finding with its consequence, remedy and file:line; the AI Act tier.

live_test_runfree

Starts the live test of a server or agent: safe calls, guarded calls (only the ones you allow), a planted instruction, leak checks. Production refused.

live_test_statusfree

Progress, the probes so far and the next task for your agent.

app_startfree

Starts a surface of your project from its code on 127.0.0.1 in a test-mode environment: the exact commands first, then your one yes.

app_statusfree

What runs where: local addresses, MCP endpoints, health, the last log lines with secrets masked.

app_stopfree

Stops everything it started, child processes too. It also stops when agentgov stops.

gov_registryfree

The registry: every agent and tool with its policy, who decided and when.

gov_policy_setGOV on

Sets a tool's policy: allowed, needs approval or forbidden, with the reason. Recorded as a decision.

gov_gateGOV on

The release gate: pass or blocked, with reasons. Every run is recorded.

gov_decision_addGOV on

Adds a decision to the register: a policy, an accepted risk, a fix, a release, a note.

gov_decisionsfree

The decision register, newest first.

gov_fix_planGOV on

The ordered fix plan from the open findings, with the remedy and the files for each. Text only; agentgov never edits your project.

gov_fix_doneGOV on

Marks a finding fixed with your evidence (the next check verifies it) or accepted with a reason.

gov_docs_exportGOV on

Writes the registry, the risk assessment, the AI Act document and the change log into .agentgov/docs/.

work_submitfree

Takes your agent's answers back and checks each one; anything wrong comes back with exactly what to fix.

agentgov_statusfree

Sign-in, licence, project folder, local data, know-how versions.

agentgov_helpfree

Tells your agent which tool fits what you said, in any language.

agentgov_sign_infree

Email sign-in in your browser, so your computer can download the know-how.

agentgov_sign_outfree

Forgets the sign-in on this computer. Your data in .agentgov/ stays.

dashboard_urlfree

Starts the private dashboard on your computer and gives you its link.

report_publishfree

Publishes this project's report at one link with an access code. The first time it only shows what would go, and asks.

report_statusfree

Every report of your account, with its link and code.

report_deletefree

Removes a report, all its versions and its code, after you confirm.

team_rulesfree

Your organisation's rules, cached on this computer.

team_requestfree

Sends a request to the admins: what, which tools, which data, why.

team_request_statusfree

The decision, its conditions and the next step (the dataready hand-off).

team_inboxfree

For admins: the requests waiting.

team_decidefree

For admins: approve, deny or approve with conditions. The member is mailed.

team_attach_reportfree

Attaches a dataready or agentgov report link to a request.

team_org_createfree

Creates your organisation; you become its admin.

team_invitefree

Invites people by mail with a join code and the install line for their tool.

team_joinfree

Joins an organisation with the code from the invite mail.

team_membersfree

Lists the members, or removes one.

team_watchfree

The MCP servers configured on this computer (names, transport, a hash; never arguments or secrets), shown first, sent to the organisation only when you say so or its rules require it.

Your data

What we see, and what we never see

Stays on your computer

Your code, your files, your keys, your prompts, the inventory, the findings, the registry and the decisions. agentgov reads your app's folder and writes only to .agentgov/ inside it. Your agent does the judging with the model you already use; we run no AI of our own.

What we keep

Your sign-in email, your organisation's rules and requests while it exists, and anonymous counts: how often a know-how pack is downloaded and how often a page of this site is opened. No cookie, no stored IP address. If you say share the report and agree, that report (counts, titles, file names with line numbers; no code or file contents) until you delete it.

What your computer downloads

The know-how: what to look for in the code, how to grade it, the AI Act questions, the safe arguments for live probes, the documentation templates. Each pack is versioned and checked before it is used, and kept so agentgov works offline.

A reading, not a verdict

Every finding is a code reading by your own coding agent, with the file and the line to check. A code reading, not a legal opinion.

The full privacy notice · Step by step