Governance checks your AI agents can call, with humans in charge.

For AI agents

Connect your agent

Read first: llms.txt and machine-interface.json list the CLI, inputs, outputs, profiles, and raw schema URLs.

Use it in your agent in 5 minutes with AI Governance MCP, the AI Governance Model Context Protocol Server. Model Context Protocol (MCP) is an open standard that lets AI assistants call external tools. It is read-only and has three tools: check_inventory, build_decision_pack, and validate_passport.

uvx --from git+https://github.com/GLOBAL-AI-GOVERNANCE/ai-governance-mcp@v0.3.1 gaig-mcp   # stdio

That one command needs only uv: the pinned Toolkit and Agentic files are bundled in the package, so nothing else is cloned or fetched. It was tested in a clean environment. To run from source instead, or to run the test suite and smoke test:

git clone https://github.com/GLOBAL-AI-GOVERNANCE/ai-governance-mcp && cd ai-governance-mcp
python3 -m venv .venv && . .venv/bin/activate && pip install -e ".[dev]"
gaig-mcp                     # stdio (Claude Desktop and other local clients)
gaig-mcp --http --port 8000  # Streamable HTTP at /mcp
pytest && python scripts/smoke_test.py   # optional checks; need the [dev] extra

On your machine

"gaig": {
  "command": "uvx",
  "args": ["--from", "git+https://github.com/GLOBAL-AI-GOVERNANCE/ai-governance-mcp@v0.3.1", "gaig-mcp"]
}

claude mcp add gaig --scope project -- \
  uvx --from git+https://github.com/GLOBAL-AI-GOVERNANCE/ai-governance-mcp@v0.3.1 gaig-mcp

grok mcp add gaig -- \
  uvx --from git+https://github.com/GLOBAL-AI-GOVERNANCE/ai-governance-mcp@v0.3.1 gaig-mcp

[mcp_servers.gaig]
command = "uvx"
args = ["--from", "git+https://github.com/GLOBAL-AI-GOVERNANCE/ai-governance-mcp@v0.3.1", "gaig-mcp"]

{"mcpServers": {"gaig": {
  "command": "uvx",
  "args": ["--from", "git+https://github.com/GLOBAL-AI-GOVERNANCE/ai-governance-mcp@v0.3.1", "gaig-mcp"]
}}}

{"servers": {"gaig": {
  "type": "stdio",
  "command": "uvx",
  "args": ["--from", "git+https://github.com/GLOBAL-AI-GOVERNANCE/ai-governance-mcp@v0.3.1", "gaig-mcp"]
}}}

Through the OpenAI, Anthropic, or xAI API

Point the provider at an HTTPS endpoint you host. Replace YOUR-HOST and TOKEN (your GAIG_MCP_BEARER_TOKEN).

{"type": "mcp",
 "server_label": "gaig",
 "server_url": "https://YOUR-HOST/mcp",
 "headers": {"Authorization": "Bearer TOKEN"}}

anthropic-beta: mcp-client-2026-09-15

"mcp_servers": [{"type": "url", "name": "gaig",
  "url": "https://YOUR-HOST/mcp",
  "authorization_token": "TOKEN"}]
"tools": [{"type": "mcp_toolset",
  "mcp_server_name": "gaig"}]

{"type": "mcp",
 "server_label": "gaig",
 "server_url": "https://YOUR-HOST/mcp",
 "headers": {"Authorization": "Bearer TOKEN"}}

Remote use needs an HTTPS endpoint you host yourself; Global AI Governance operates no hosted endpoint. Any network-reachable server must set GAIG_MCP_ALLOWED_HOSTS and GAIG_MCP_BEARER_TOKEN (32+ characters); without them it refuses to start. Consumer connector UIs that cannot send a static bearer token (ChatGPT on the web today; Claude.ai outside its request-headers beta) cannot connect. These remote recipes have not been tested against a live provider; see the repository README. The command is gaig-mcp and the Python package is gaig_mcp.

Where your data goes: the browser Toolkit processes files only in this browser. Through MCP, the inventory or passport you pass becomes part of your AI client's conversation, so a hosted assistant sends it to its provider, and a remote server you host receives it over HTTPS. validate_passport checks a passport document; it does not monitor, enforce, or revoke anything at runtime.

For agent builders

Guard every tool call

Your agent has more tools than these three. A small Action Level guard (Python standard library) checks each tool call before it runs, using the Action Levels in the action-authority spec: Levels 0 to 3 run, Level 4 waits for an accountable person to approve or reject it, Level 5 is blocked, and a tool missing from your policy counts as Level 4. The guard never approves a Level 4 call itself.

Files and setup: AI Governance MCP README (release v0.3.1). The guard limits what an agent can do unattended; it does not make any action authorized.

Specifications

Read the spec

• Global AI Assurance Envelope: the portfolio-level, reference-only interoperability object.
• Toolkit machine interface and v2.7.1 contracts.
• Agentic AI Governance v0.1.0-alpha.2 schemas: agent authority, passport, expiry, and revocation.

Product map

What exists today