Connect AgentBee to your AI
One step per tool. After that, you just tell your agent in plain English which actions to gate, and the bee does the rest.
The one rule: AgentBee is a USB device, so it gates where your agent's actions actually run, on your machine. Any local agent works (Claude, Codex, Gemini, an Agents-SDK app). A cloud chat window in a browser cannot reach a USB key on your desk, and neither can anything else, that is the whole point. Connect it to the runtime, not the chat box.
Find your bee's port first: ls /dev/cu.usbmodem* on macOS, or ls /dev/ttyACM* on Linux. Use it as AGENTBEE_PORT below.
Claude (Code / Desktop) MCP
One command, then just ask Claude to gate things.
claude mcp add agentbee \ --env AGENTBEE_MODE=usb --env AGENTBEE_PORT=/dev/cu.usbmodemXXX \ -- npx -y @agentbee/mcp
Then say: "Gate my critical actions with AgentBee."
OpenAI Codex CLI MCP
Append the server to your config (paste-safe, one line):
printf '\n[mcp_servers.agentbee]\ncommand = "npx"\nargs = ["-y", "@agentbee/mcp"]\nenv = { AGENTBEE_MODE = "usb", AGENTBEE_PORT = "/dev/cu.usbmodemXXX" }\n' >> ~/.codex/config.toml
codex mcp list # confirm 'agentbee' is listed
codex
Then ask: "Use the agentbee request_approval tool before any destructive action."
Gemini CLI MCP
gemini mcp add -s user --trust \ -e AGENTBEE_MODE=usb -e AGENTBEE_PORT=/dev/cu.usbmodemXXX \ agentbee npx -y @agentbee/mcp gemini mcp list
Any MCP client (OpenAI Agents, Cursor, Zed, …) MCP
Add this server to the client's MCP config:
{
"mcpServers": {
"agentbee": {
"command": "npx",
"args": ["-y", "@agentbee/mcp"],
"env": { "AGENTBEE_MODE": "usb", "AGENTBEE_PORT": "/dev/cu.usbmodemXXX" }
}
}
}
No MCP? (Aider, scripts, CI) HOOK
Gate at the action instead. A git pre-push hook so pushing to main needs a hold:
npx -y @agentbee/cli install-hook pre-push # drops the gate + hook into .git/hooks
Or in your own code: npm i @agentbee/sdk and call await bee.approve({action, scope, trust}) before the risky step.
Two things worth knowing
- AGENTBEE_MODE=usb is required. Without it the server fails closed and will not auto-approve. Simulator mode (auto-approve, no human) is dev-only and must be turned on with
AGENTBEE_ALLOW_SIMULATOR=1. - Cooperative vs enforced. The MCP
request_approvaltool relies on the model choosing to call it. For un-bypassable gating, use the hook path (or a PreToolUse hook): it intercepts the action whether or not the model cooperates.