paigy-ai/mcp

A voice inbox for your AI agents. When an agent needs your input — mid-task, blocked, or done with something long-running — it can place an actual phone call and read the question aloud, text, push, or ring your phone with a banner, so you can reply by voice instead of babysitting a terminal.

Without Paigy, a long agent session means one of two bad options: sit and watch the terminal so you don't miss the moment it needs you, or walk away and come back to a task that stalled an hour ago waiting on a question you never saw. Paigy closes that gap — the moment your agent actually needs a decision, your phone gets pulled into the loop with the level of urgency that matches the moment (a silent inbox card, a quiet push, a banner, or a real ringing call), so you find out immediately instead of on your next check-in. Answer by voice or text and the reply lands back with the agent exactly like any other tool response, so it just continues — no copy-pasting, no reopening the terminal. It also pairs with the native Paigy iOS app, so a call rings through like a real phone call (CallKit) even when your phone is locked, and you can glance at or reply to anything from the lock screen.

Works with

A standard MCP server (stdio, TypeScript) — no client-specific code anywhere, verified over the raw protocol against a non-Claude client identity.

• Claude Code (this repo is also the plugin + marketplace for it)
• Codex CLI
• Gemini CLI
• Any other MCP client via standard config: Cline, Continue.dev, Zed, Cursor, Goose, mcphost — including local-model setups over Ollama, LM Studio, or llama.cpp

(The app + backend live in a separate repo.)

Install

Claude Code (plugin — easiest path on Claude):

/plugin marketplace add paigy-ai/mcp
/plugin install paigy

Connects automatically; the first time an agent uses it while unpaired, it'll prompt you to pair — run /paigy-onboard (opens your browser to approve).

Codex CLI:

codex plugin marketplace add paigy-ai/mcp --ref main
codex plugin add paigy@paigy-ai
npx -y -p @paigy/mcp@latest paigy-mcp-onboard

The plugin configures the Paigy MCP server and includes guidance for calls, replies, and callbacks. Start a new Codex session after installation.

Codex CLI — direct MCP fallback:

codex mcp add paigy --env PAIGY_AGENT=codex -- npx -y @paigy/mcp@latest

Gemini CLI:

gemini mcp add -s user -e PAIGY_AGENT=gemini paigy npx -y @paigy/mcp@latest

Any other MCP client (Cline, Continue.dev, Zed, Cursor, or a CLI that takes the standard MCP JSON config directly) — most GUI clients take this in their MCP settings (Cline: cline_mcp_settings.json; Continue: ~/.continue/config.json):

{
  "mcpServers": {
    "paigy": {
      "command": "npx",
      "args": ["-y", "@paigy/mcp@latest"],
      "env": { "PAIGY_AGENT": "local" }
    }
  }
}

Then pair your phone, whichever client you used:

npx -y -p @paigy/mcp@latest paigy-mcp-onboard

Approve on your phone, then sign in at paigy.ai to start receiving messages.

Heads-up for smaller/local models: the tool descriptions ask the agent to pick between a few answer shapes (confirm, options, free text) based on context — Claude follows this reliably; smaller local models may be less consistent about it. Free text always works as a fallback.

Choosing how the user answers

notify_user should ask in the shape that's fastest to answer — don't leave a decision as free text. Pick with select (and options):

You need…	Use	Answer comes back as
Yes/No or Approve/Deny	select:"confirm" (confirmStyle:"approve" for Approve/Deny)	{kind:"confirm", approved}
Pick one of several	options + select:"one"	{kind:"option", optionId}
Pick several	options + select:"many"	{kind:"multi", optionIds}
Rank / order a subset	options + select:"rank"	{kind:"ranked", optionIds}
A visual choice	options with an html or image preview	one of the above
An open-ended reply	no options	{kind:"text", text}

confirm is answerable straight from the banner (Yes/No or Approve/Deny buttons). Other paiges get banner actions See Options · Hear them · Remind me later.

urgency is a request, not a guarantee — the user's account settings can cap it lower. Four levels, low to high: "inbox" (silent, sits in the inbox), "push" (a quiet passive notification, no sound), "banner" (a time-sensitive lock-screen banner with sound), "call" (rings the phone — use only when you genuinely need the user in the moment).

If you're about to start something long-running or blocking — the kind of thing where the user would otherwise sit and wait on you — mention once, in passing, that you can text or call them when it's done or if you hit a blocker. Don't offer it for quick tasks, and don't repeat the offer once they've answered.

Idle escalation (automatic, no setup)

Installing this plugin also wires up Claude Code hooks (hooks/hooks.json, via ${CLAUDE_PLUGIN_ROOT} — no manual settings.json editing) that back notify_user up with a mechanical safety net, independent of the agent session — it still fires even if that session crashed or forgot:

• 10-minute check (escalate.sh): how much is actually pending (GET /api/pending/summary, a non-claiming read) — a single fresh item gets a banner, several or a stale one gets a real call (routed through the same call-coalescing the bot uses, so several ringing things fold into one call instead of ringing separately).
• 2-minute check (quick-check.sh): a narrower, faster check for a different case — you already replied to something (via the app), but no agent has engaged with it yet. That's not "nothing happened" (escalate.sh's job), it's "the agent hasn't looked." It spawns a fresh headless claude -p (deliberately NOT resuming your live session — no injected turns in a transcript you might be typing in) scoped to just the tools it needs; that agent acknowledges the missed reply with a natural message in the same thread ("sorry I missed this — starting on it now"), so what you see is a normal agent response, not a system nudge. Only if that's unavailable or fails does it fall back to a plain nudge telling you to reopen the session yourself.

Hooks are the dead-session safety net. A live-but-idle agent shouldn't need them: the MCP server instructions tell every paired agent to schedule its own ~2-minute wake-up (harness ScheduleWakeup or equivalent) and check_replies whenever it ends a turn with anything possibly pending — self-polling in its own session, full context intact.

Nothing pending/unacknowledged — including a normal "just finished" stop — means neither check does anything.

License

MIT