ringback

Your AI agent can call your phone — and actually talk to you.

ringback gives an LLM (Claude, or any MCP client) tools to reach you on your phone — from a one-way "fierce" alert all the way to a live, interruptible voice conversation — using only free, self-hosted pieces. No paid telephony. No extra API key for the conversation: the model already driving the MCP is the voice on the line.

Highlights

• 📞 Two-way voice calls — the agent rings your phone, you talk, it transcribes you and replies in speech. Barge-in: talk over it and it stops.

• 🔔 Tiered alerts — a loud push (ntfy / Pushover) or a real SIP ring + chat message, fired only when the LLM judges it urgent.

• 🆓 Free & self-hosted — pjsua2 + whisper.cpp + macOS say + a free Linphone SIP account. No Twilio, no per-minute fees.

• 🧠 No conversation API key — the calling model is the brain; these tools are just its ears and mouth.

It ships two MCP servers, ringback-alert and ringback-voice:

Platform: macOS only. The voice feature uses Apple's say (TTS) and CoreAudio via pjsua2.

The criteria for when to contact you live in the tool descriptions — the calling LLM decides. These servers are just the mechanism.

What a call looks like

agent → call_start("Your nightly deploy failed — want me to walk you through it?")
   📞 your phone rings; you pick up and hear the line
you   → "yeah, which step broke?"
agent → "The database migration. I can roll it back and retry — want that?"
you   → "yes, do it"            ← you can also just talk over the agent to interrupt
agent → call_end()

The LLM calls call_start once, then converse(...) for each turn. Plain alerts are even simpler: one alert_me(...) call.

Quick start

git clone https://github.com/mohitbadwal/ringback && cd ringback
./setup.sh        # installs EVERYTHING (toolchain, pjsua2, whisper model, deps) + creates voice.env
# edit voice.env → add your 3 SIP vars (free account: https://subscribe.linphone.org), then:
claude mcp add ringback-voice --scope user -- "$PWD/run_voice_mcp.sh"

Full walkthrough + env-var reference: Set up ringback-voice below.

Honest caveats (read first)

• macOS only. say + CoreAudio. No Linux/Windows path today.

• Not ChatGPT-realtime. The voice loop is record → whisper STT → LLM → say TTS, so expect ~1–2 s per turn. It's a reliable walkie-talkie with barge-in, not a streaming realtime voice.

• The voice feature depends on GPL software (pjproject/pjsua2). This repo is Apache-2.0, but redistributing a bundle that links pjsua2 carries GPL obligations — see NOTICE. The ringback-alert server is unaffected.

• Your Mac must be awake and online, and for a voice call a Claude session must be live (it's the brain) for the duration.

• Barge-in assumes low acoustic echo (handset or headset). On speakerphone, the TTS can echo into the mic and false-trigger "interruption." There's no echo cancellation in this path.

• iOS push reality: a self-hosted/free push can't truly pierce Focus/Silent on iPhone except via Pushover's Critical Alerts (paid) — see ringback-alert notes below.

Architecture

  LLM (Claude)  ──MCP tools──▶  ringback-voice server (Python)
                                   │  call_start / converse / listen / speak
                                   ▼
                 pjsua2 (SIP+SRTP, built from source)  ──▶  Linphone SIP server
                   │  say → ffmpeg → WAV  (speak)              │ APNs VoIP push
                   │  record → whisper.cpp (listen)            ▼
                   └───────────────────────────────────▶  your iPhone rings

ringback-alert is simpler: it shells out to ntfy/Pushover HTTP and/or baresip for a SIP ring + chat message.

Prerequisites

• macOS (Apple Silicon or Intel) with Homebrew

• A free Linphone SIP account (sip.linphone.org) and the Linphone iOS app (for the ring/voice features)

• Python 3.10+ (the pjsua2 bindings are built against whichever python3 you point at)

Set up ringback-voice — 4 steps

1. Clone + install everything:

git clone https://github.com/mohitbadwal/ringback && cd ringback
./setup.sh

setup.sh installs the toolchain, compiles pjsua2 from source (~20–30 min — no Homebrew formula exists for the bindings), downloads the whisper model, installs deps, and creates voice.env for you. Safe to re-run. (Override PYTHON_BIN / PJPROJECT_DIR / WHISPER_MODEL_NAME if your layout differs.)

2. Get a free SIP account (this is the phone that rings):

• Sign up at https://subscribe.linphone.org (or tap Create account in the Linphone app). You get a username and password; your address is sip:<username>@sip.linphone.org.

• Install the Linphone app on your iPhone, sign in, and confirm it shows Connected.

3. Fill in voice.env (already created by setup.sh — just edit it). Only three vars are required:

export VOICE_SIP_ID="sip:yourname@sip.linphone.org"
export VOICE_SIP_USER="yourname"
export VOICE_SIP_PASS="your-password"

Full variable reference:

4. Register + test:

claude mcp add ringback-voice --scope user -- "$PWD/run_voice_mcp.sh"

Then in a fresh Claude session say: "Use ringback-voice to call me and say hello." Your phone should ring.

Claude Desktop instead of Code? Add this to ~/Library/Application Support/Claude/claude_desktop_config.json (absolute path required; restart the app):

{ "mcpServers": { "ringback-voice": { "command": "/absolute/path/to/ringback/run_voice_mcp.sh" } } }

Set up ringback-alert (optional)

ringback-alert reads its config from the MCP client's env block (no file to source). Register it with the channels you want:

# Claude Code
claude mcp add ringback-alert --scope user \
  --env ALERT_CHANNEL=ntfy \
  --env NTFY_URL=https://ntfy.sh/your-long-random-topic \
  -- /opt/homebrew/bin/uv --directory "$PWD" run server.py

See alert.env.example for all variables (ntfy / Pushover / SIP ring). Use a long random ntfy topic — anyone who knows it can read/publish.

Using ringback-voice (the conversation)

The LLM drives a simple loop:

reply = call_start("Hi, it's your assistant — your deploy failed. Want details?")
# rings the phone, speaks the line, returns the user's first words
reply = converse("It failed on the database migration step. Want me to retry it?")
# speaks AND listens in one interruptible turn
... repeat converse() each turn ...
call_end()   # when the user says "bye" / hangs up

• converse(text) speaks while listening. If you talk over it, it stops immediately and tells the LLM how far it got and what you said (barge-in).

• get_conversation() returns the full transcript so far — both sides, plus where it got interrupted.

• TTS reads text literally, so the tool descriptions instruct the model to speak plain-language summaries, never raw logs/codes — those go via alert_me as text instead.

Whisper model accuracy/speed trade-off (set WHISPER_MODEL): base.en (fast/rough) → small.en (balanced, default) → medium.en (most accurate/slow).

Using ringback-alert (notifications)

alert_me(message, severity, title) with severity = info | warn | critical. Channels via ALERT_CHANNEL (comma list of ntfy, pushover, call):

• ntfy — free push; loud but does not pierce iOS Focus/Silent unless whitelisted per-Focus.

• Pushover — $5 one-time; true iOS Critical Alerts (pierces Focus/Silent, repeats until acknowledged) at critical.

• call — free SIP ring + Linphone chat message via baresip; rings full-screen, only at critical by default.

A built-in rate-limit guard (default 5/60s, per-process) stops a misfiring caller from spamming you.

Bundled skill: watchdog

skills/watchdog/ is a ready-to-use Claude skill built on these servers. It watches a condition you give it (a CI run, a deploy, a pod, a file) and escalates only when you're actually away from the laptop — chat status → chat warning → ringback-alert push → ringback-voice call — judged by macOS HID idle time. It never interrupts you while you're typing, and de-escalates the moment you touch the keyboard.

cp -r skills/watchdog ~/.claude/skills/watchdog   # install for Claude Code

Then: /watchdog <what to watch> | priority=<low|medium|critical> — low = chat only, medium = may send a phone alert, critical = may place a call. Full design in skills/watchdog/SKILL.md.

Security

• SIP credentials live only in your local, gitignored voice.env (and the baresip accounts file for ringback-alert) — never in the repo or the MCP client config when you can avoid it.

• The voice server only ever calls the single SIP URI you configure — it cannot dial arbitrary numbers.

• Treat ntfy topics as secrets (use a long random topic); don't put sensitive detail in alert bodies on public ntfy.sh.

• See NOTICE for the GPL/pjproject licensing caveat before redistributing.

License

Apache-2.0 (see LICENSE), with an important GPL caveat for the voice component's pjproject dependency — see NOTICE.

Credits

Built on pjproject/pjsua2, whisper.cpp, baresip, ntfy, and Linphone.