A deterministic policy firewall for AI agent tool calls. YAML rules, no LLM calls, no risk scores. CLI + MCP server (advisory) + @enforce() decorator (unbypassable). MIT, 46 tests, zero network dependency.
Setup
Choose your client, copy the config, then fill in any required environment variables.
claude mcp add agent-guardrail -- uvx guardrail-mcp~/.codex/config.toml
[mcp_servers.agent-guardrail]
command = "uvx"
args = ["guardrail-mcp"]
# codex mcp add agent-guardrail -- uvx guardrail-mcp~/.cursor/mcp.json
{
"mcpServers": {
"agent-guardrail": {
"command": "uvx",
"args": [
"guardrail-mcp"
]
}
}
}.vscode/mcp.json
{
"servers": {
"agent-guardrail": {
"type": "stdio",
"command": "uvx",
"args": [
"guardrail-mcp"
]
}
}
}~/.gemini/settings.json
{
"mcpServers": {
"agent-guardrail": {
"command": "uvx",
"args": [
"guardrail-mcp"
]
}
}
}claude_desktop_config.json
{
"mcpServers": {
"agent-guardrail": {
"command": "uvx",
"args": [
"guardrail-mcp"
]
}
}
}.trae/mcp.json
{
"mcpServers": {
"agent-guardrail": {
"command": "uvx",
"args": [
"guardrail-mcp"
]
}
}
}