Agent setup
Set up Kilobat in Codex CLI
Codex can check the network before a big download, find why its own request hangs and keep the Mac awake for a long task. One line in config.toml tells Kilobat when a turn is done.
Install the CLI
The Kilobat CLI is free and open source, and it includes the MCP server. Install it with Homebrew:
brew install seifip/kilobat/kilobatCheck it with kilobat --version. kilobat doctor says whether the Kilobat app is running, its network filter is on and Pro is active; the tools work without the app, which adds Wi-Fi names, history and rule enforcement. The CLI needs macOS 14 or later.
Connect the MCP server
Add the server from the terminal:
codex mcp add kilobat -- kilobat mcpOr add it to
~/.codex/config.tomlyourself. A longer tool timeout leaves room for speed and stutter tests.[mcp_servers.kilobat] command = "kilobat" args = ["mcp"] tool_timeout_sec = 120
A .codex/config.toml in a trusted project adds the server for that project only.
Add the skills
Nine Agent Skills teach the agent to read Kilobat's measurements and to ask before changing settings. Install them for all your projects with skills.sh:
npx skills add seifip/kilobat-skills -g -a codex
This puts them in ~/.agents/skills/. Leave out -g to install them for the current project only, or leave out both flags to pick agents from a list.
The kilobat skill is the entry point: it holds the setup and safety rules and sends the agent to the skill that fits, such as mac-network-troubleshooting or mac-agent-runs. See all nine skills.
Try these prompts
Ask in plain words. The agent picks the tools; these are the ones each request usually needs.
| Ask | MCP tools |
|---|---|
| npm install hangs on registry.npmjs.org. Find the layer that fails. | diagnose_connection |
| The internet feels slow. Find out why and tell me what to fix first. | diagnose_network, get_status |
| I'm about to pull a 6 GB Docker image. Is now a good time, or am I on a metered network? | get_conditions |
| Keep the Mac awake for the next two hours while you run the full test suite, then release it. | keep_awake, release_awake |
| Trace the route to github.com three times and tell me where the delay starts. | trace_route, get_route_history |
| Did my agent runs today finish, or did one stall or lose the network? | list_sessions |
What Codex CLI can and can't change
Kilobat's MCP server marks every tool as read-only or as a change, and each change tool's description tells the agent to get your confirmation first. The kilobat skill holds the same rule: read freely, ask before changing.
Runs without asking
- Measure Wi-Fi, router, internet and DNS latency, and per-app traffic:
diagnose_network,get_status,get_wifi,scan_wifi. - Find the layer that stops one host, trace a route and compare resolvers:
diagnose_connection,trace_route,benchmark_dns. - Read history, sessions, rules and approvals:
get_usage_report,list_sessions,list_rules,list_approvals. - Ping for latency spikes with
run_stutter_test, which sends pings only.
Asks you first
- A speed test, which uses up to 80 MB (
run_speedtest; on a metered network, 10 MB or less), and a five-second LAN test (run_lan_test). - Looking up your public IP with Cloudflare (
get_exit_pathwithpublic_ip). - Profiles, app rules, blocked hosts, the new-app mode, budgets, quotas, a network's data cost and pausing updaters (
set_profile,set_app_rule,block_destination,set_new_app_mode,set_budget,set_app_quota,set_network_cost,set_updater_pause). - DNS, which macOS also guards with an administrator password prompt (
set_dns). - Keeping the Mac running with the lid closed, every time (
keep_awakewithlid_closed). A plainkeep_awakechanges no settings and ends when it expires.
Can't do
- Let a waiting app connect on its own:
decide_approvalcan block at once, but allowing opens Kilobat's approval window for you to confirm. - Change AirDrop visibility:
set_airdrop_visibilityopens System Settings for you. - Change sleep settings, another app's settings, or anything outside Kilobat. It never reads prompts or replies.
- Enforce rules without the Kilobat app and its network filter (Pro). Without them, rules are saved, not applied, and the agent should say so.
Codex shows Kilobat's tools with their annotations; keep its approval prompts on for the change tools.
If the agent can also run shell commands, the kilobat CLI can do what the MCP tools do. The skills tell it which commands to run freely and which to ask about.
Track agent runs
Simplest: notify in ~/.codex/config.toml. Codex passes a JSON argument after each turn, which --json-arg reads. It only says done, and Codex runs one notify program.
notify = ["kilobat", "agent-event", "done", "--agent", "codex", "--json-arg"]
For waiting too, use Codex hooks in ~/.codex/hooks.json instead, and trust them with /hooks:
{
"hooks": {
"SessionStart": [{ "hooks": [{ "type": "command", "command": "kilobat agent-event --agent codex --stdin-json" }] }],
"UserPromptSubmit": [{ "hooks": [{ "type": "command", "command": "kilobat agent-event --agent codex --stdin-json" }] }],
"PermissionRequest": [{ "hooks": [{ "type": "command", "command": "kilobat agent-event --agent codex --stdin-json" }] }],
"Stop": [{ "hooks": [{ "type": "command", "command": "kilobat agent-event --agent codex --stdin-json" }] }]
}
}Sources
- OpenAI Codex docs: Model Context Protocol
- Kilobat docs: AI agents and MCP
- skills.sh: supported agents and their skill folders
Checked on 30 October 2026. Agents change their settings often; if a step no longer matches, tell us at the support address.