Pi coding agent

Human-in-the-loop for pi, in one command.

Pi is a terminal coding agent. It has no built-in MCP support, so ping-a-human-pi ships a tiny embedded MCP client that connects ping-a-human to pi's hooks.

Install

npx ping-a-human-pi

This single command:

installs a pi extension into ~/.pi/agent/extensions/ping-a-human-pi/,
installs the pah CLI onto your PATH,
seeds ~/.pi/agent/ping-a-human.json with defaults (never overwrites an existing one).

Run /reload in pi (or restart it). Remove everything with npx ping-a-human-pi uninstall.

First time with ping-a-human? Configure Telegram once with npx ping-a-human setup.

What you get

Tools — notify_human and ask_human become native pi tools the model can call.
Notifications — a Telegram ping with a short summary when each task finishes (pi's agent_end hook).
Approval — before risky tool calls (pi's tool_call hook), pi asks via ask_human and blocks unless you approve. Off by default.
/ping command — inside pi: /ping or /ping your message.
pah CLI — from any terminal, independent of pi.

The pah CLI

pah ping "build finished"      # fire-and-forget notification
pah notify "deploy started"    # alias for ping
pah ask "ship to prod?"        # waits and prints the human's reply
pah help

Because pah ask prints the reply to stdout, you can branch on it in scripts:

if pah ask "deploy to prod?" | grep -qiE '^(y|yes|ok|approve)'; then
  ./deploy.sh
fi

Configuration

All optional — edit ~/.pi/agent/ping-a-human.json:

Key	Default	Meaning
`command` / `args`	`npx -y ping-a-human`	how to launch the server
`exposeTools`	`true`	register `notify_human` / `ask_human` as model-callable tools
`notify.enabled`	`true`	ping you when a task finishes
`notify.template`	`✅ pi finished a task in {cwd}…`	message text; `{summary}` `{cwd}` substituted
`approval.enabled`	`false`	ask before risky tool calls
`approval.tools`	`["bash"]`	which pi tools to gate
`approval.patterns`	danger regexes	only gate calls whose args match one; empty = gate all

 
A reply matching yes/ok/approve/allow/sure/go/lgtm/… approves;
    anything else blocks the call, and the reason is sent back to the model.
  ← End-to-end verification