Changelog
What's shipping and what just changed.
We post release notes here whenever something affects how calls are placed, how agents should use Ring-a-Ding, or how the service behaves. Newest first.
Subscribers don't have to do anything to pick up most changes — the API server and the docs update themselves. CLI updates ship via npm: run npm install -g ring-a-ding-cli@latest to pull the newest version.
July 16, 2026
Sharper, smoother calls with GPT-Realtime-2.1
Calls now run on OpenAI's GPT-Realtime-2.1, with better recognition of letters and numbers, steadier handling of silence and background noise, and more natural interruptions — at the same audio-token rates as GPT-Realtime-2.
- —Every outbound call now uses gpt-realtime-2.1 through the latest OpenAI Agents SDK.
- —Confirmation codes, phone numbers, email addresses, and other spelled-out details are easier for the agent to recognize accurately.
- —Pauses, noisy lines, and barge-in interruptions are handled more reliably using the model's automatic turn-detection behavior.
- —No action required. Existing CLI and MCP installations receive the upgrade automatically.
June 2, 2026
Where Ring-a-Ding can call: US & Canada only, now documented
Ring-a-Ding places calls to US and Canadian numbers only — international calling isn't supported, and the outbound number pool is US-based. That was always true at the telephony layer; now it's stated plainly across the site, docs, terms, and the agent skill so nobody finds out mid-call.
- —Added a "Which countries can I call?" FAQ on the homepage and a US-and-Canada-only note in the docs limits and troubleshooting.
- —Terms of Service now states calls are limited to US and Canada destinations.
- —The OpenClaw skill and the make_call MCP tool description now say US & Canada only, so your agent knows before it dials and can tell you up front.
- —No change to how calls work — international numbers were already rejected by the carrier; this just makes the limitation clear ahead of time.
May 11, 2026
Calls now run on gpt-realtime-2
We've upgraded the underlying voice model to OpenAI's new gpt-realtime-2 — GPT-5-class reasoning, a 128K context window, and a 26-point lift in call success on hard adversarial scenarios (95% vs 69%). Same cost on your BYOK OpenAI key; better calls.
- —Default model on every outbound call switched from gpt-realtime-1.5 to gpt-realtime-2.
- —No price change on audio tokens. Same $32 / 1M audio input and $64 / 1M audio output on your BYOK OpenAI key.
- —Stronger instruction following and more reliable tool use — the agent is meaningfully better at navigating phone trees, holding longer turns, and ending calls cleanly when the conversation wraps.
- —No action required. Your existing CLI / MCP setup picks this up automatically.
April 23, 2026
Fixed an outage that made every outbound call fail silently
Outbound calls via `POST /calls/make` were returning bare 500s for about a week. Two Supabase migrations that the API requires (a `brief_notes` column and a TEXT→JSONB shape change on `summary`) had never been applied to production. Both are now applied; calls are working end-to-end.
- —Applied migration `004_brief_notes` — adds the `brief_notes JSONB` column the API inserts on every call.
- —Applied migration `005_structured_summary` — converts the `calls.summary` column from TEXT to JSONB so the post-call structured-summary write succeeds.
- —The failure mode was ugly: Postgres rejected the INSERT, the Supabase error object broke pino serialization inside the Hono `onError` handler, so the app never wrote a response and Fly's proxy returned a bare `500 text/plain` with no log line. If you called `/calls/make` and got a mysterious 500, this was why.
- —No code changes or redeploys were needed — the migrations brought prod's schema back in sync with the deployed API.
April 22, 2026
Keyword-rich skill listing, unified header mark
Ring-a-Ding's public ClawHub listing gets a full rewrite — the description and README now reflect what the skill is actually for, including the hands-free call-your-user use case. The header logo mark now matches the favicon.
- —ClawHub listing copy rewritten: new tagline highlights appointments, reservations, voicemail, and hands-free agent-calls-you conversations. README body leads with a rich opening paragraph and an "Example use cases" section.
- —npm keywords expanded from 6 items to ~40 — covers phone/telephone/cell-phone/voicemail, voice-ai/voice-agent, openai-realtime/gpt-realtime/twilio, mcp/openclaw/claude, appointment-booking/scheduling/reservation/hands-free, and more. Surfaces on `npm search` and third-party indexers.
- —The website header and footer "R" mark now renders from the same SVG as the favicon — one source of truth for the logo across every surface.
- —CLI 0.2.1 released to npm.
April 22, 2026
Marin default, fair number rotation, agent onboarding at install time
The default voice is marin again, outbound phone numbers rotate evenly across the pool, and AI agents receive an identity-forward onboarding briefing the moment they install the CLI.
- —Default voice for new calls is now marin (was silently defaulting to alloy due to a database fallback mismatch). All existing accounts were reset to marin.
- —Outbound calls now pick phone numbers at random from the pool instead of hitting the same handful over and over. Usage spreads evenly across every number in the pool.
- —New agent briefing at install time: the output of `npm install -g ring-a-ding-cli` and the new `rad briefing` command speak directly to the AI agent, instructing it to proactively offer a welcome test call on first use and compose a rich, identity-forward brief (its own voice, caller name, personality, purpose, context, and opening line).
- —The OpenClaw install snippet is now four steps: `npm install`, `openclaw skills install`, `openclaw skills check`, `rad briefing`. The last step prints the agent briefing and a quick install-status summary.
- —SKILL.md (what OpenClaw loads into its context) gets a new "First-time use" section reinforcing the same proactive behavior from inside the agent loop.
- —Website readability pass: body text is slightly larger and higher-contrast across every page — easier to read without losing the dark-luxury look.
- —Site now has a proper favicon (the amber R mark) and iOS home-screen icon.
- —CLI 0.2.0 released to npm.
April 16, 2026
Structured call summaries, agent cron polling, brief quality notes
Every completed call now returns a structured summary you can act on without re-reading the transcript, and agents can poll for call results without blocking themselves.
- —Every completed call returns a structured `summary` object: `headline`, `decisions[]`, `openQuestions[]`, and up to five keeper `quotes[]` (each tagged as the assistant or the other party).
- —`make_call` now returns an `agent_followup` block with ready-to-execute `cron_add` and `cron_remove` payloads. OpenClaw and similar orchestrators can register a single cron job to poll the call result, then auto-remove it when the call terminates — no more blocking the entire agent for up to 31 minutes.
- —`briefNotes[]` gives you quality feedback about the brief you sent in (e.g. "no personality was provided, default used"). Read these between calls to improve the next brief.
April 14, 2026
Voice guide expanded, homepage polish
The voice guide now describes gender, character, and ideal use case for all ten available voices, and the marketing site got a small refresh.
- —Voice guide in /docs covers all ten voices (marin, alloy, ash, ballad, cedar, coral, echo, sage, shimmer, verse) with gender, character, and best-fit use case.
- —Start with marin (female, natural, default). Pair voices with `--personality` for best results.
- —Homepage hero copy refreshed and a copy-to-clipboard install box added for quick onboarding.