Callwright

Callwright

A voice agent that places phone calls — reservations, appointments, confirmations, general inquiries — on your behalf, shaped by an LLM's direction + grounding. It discloses it's an AI, negotiates within bounds you set, never fabricates, and reports the outcome.

Built on Retell. Exposed as an MCP server so any MCP-capable assistant (Claude, ChatGPT, Cursor, Copilot) can place calls for you — combining callwright (the actuator) with your other MCPs (WorkIQ, Gmail, calendar — the sensors) for context.

Just ask

You don't fill out forms — you ask your assistant in plain language, and it gathers the context (from your calendar, email, or the chat), shapes the call, reads it back to you, and dials on your OK:

"Find the repair appointment in my calendar and call the contractor to confirm it's still on for Thursday."

"Call Sakura Sushi and book a table for 4 this Friday at 7 — anytime between 6:30 and 8 is fine."

"Ring my dentist and move my cleaning to sometime next week; my member ID is in my profile."

"電話して、東京の美容院に明日の午後カットの予約を取って." — (places the call in Japanese)

"Call this hotel in Kyoto and ask their hire-car rate to the airport — just gathering info, don't book."

The assistant pulls only the minimal details each call needs, confirms the plan with you first, then places it fire-and-forget and fetches the outcome when you ask.

How it works

You (chat) --> Host LLM --gathers context from other MCPs--> shapes a place_call                  |                  v            callwright MCP  --validate + profile + compose + guard--> Retell --> phone                  |                                                     |                  +-- get_call_outcome / learn_from_call <-- outcome --+

• Generic agent, not per-scenario. One agent prompt = fixed guardrails (disclose AI, never pay, never fabricate, PII-minimize) + injected per-call direction. Handles any scenario and language.
• Multilingual, extensible at runtime. Ships speaking English and Japanese; add_language stands up a brand-new call language (translated prompt + a native voice + its own agent) entirely from chat — no redeploy. All language phrasing is data (lang-phrases.json), not code.
• Purpose-first disclosure. Opens with the goal ("I'm an AI assistant. I'm calling to confirm..."), not a name. Your name surfaces only when a business needs it.
• Data minimization, filled lazily. PII lives in principal.facts (shared only if asked); each call carries only the minimal relevant subset, and the read-back lists exactly which data types go out. When a call uses a durable fact you haven't saved, it offers to remember it — so the profile fills from use, not an upfront interview.
• Learns from every call. When a call defers on a missing detail, the profile is enriched so next time it's collected up front — manually (learn_from_call) or automatically via a signature-verified Retell webhook. A genuinely new scenario is staged and proposed (never silently created), with guards against over-specific splintering.
• Bounded retry. An unanswered/busy call is re-dialed — exactly once by default, never more without an explicit request (anti-spam).

Tools

Setup & profiles

Calling

Languages

Learning

place_call is fire-and-forget and gated: the first call returns a read-back (including the PII footprint); you dial only after confirming with confirm:true.

get_call_outcome returns a canonical outcome object. Its custom-analysis-first status is authoritative when Retell's built-in call_successful signal disagrees; both raw signals remain available for diagnosis. Successful informational calls use completed, while booked is reserved for an actual reservation, appointment, or commitment.

Retell setup (prerequisites)

Callwright dials through Retell, so you need a Retell account with outbound calling enabled before anything works. In the Retell dashboard:

1. Create an account and add a payment method (Stripe). Calls and phone numbers are pay-as-you-go; there's no Callwright fee on top.
2. Complete KYC verification. This is required to unlock outbound calling and phone-number purchases — without it, calls are blocked. (It's a one-time identity/business check; approval is usually quick.)
3. Buy a phone number. This becomes your caller ID (the from number). A US local number is cheapest; Callwright auto-selects it during configure run_provisioning.
4. Create an API key (Settings → API Keys). This is your RETELL_API_KEY — it stays server-side and must never go in a browser or chat.

Optional, recommended:

• Verified phone number / branded calling (dashboard → Telephony) — registers your number so carriers are less likely to tag it "Spam Likely", which meaningfully improves pickup rates on outbound calls.

Not required (and currently dormant here):

• A2P 10DLC registration — only needed to send SMS, via a multi-step business-profile → brand → campaign approval with a recurring cost. Callwright's SMS notifications are built but shelved, so you can skip this entirely for placing calls.

Costs are per-minute and destination-driven (roughly ~$0.015/min US, higher internationally) plus the monthly number rental.

Install

Published on npm as callwright. Install globally to get the callwright (server) and callwright-init (setup wizard) commands:

npm install -g callwright

Or clone this repo and npm install to run from source (node server.js, node init.js).

Either way, callwright is a self-hosted, single-user HTTP MCP server: you run your own instance with your own Retell credentials (see Retell setup above). Nothing is shared or multi-tenant.

Run locally

Using the global install:

callwright-init                     # one-time setup wizard (number, agent, your identity)export RETELL_API_KEY=key_xxx       # [Environment]::SetEnvironmentVariable on Windowscallwright                          # MCP server on :8787 (POST /mcp)

From a cloned repo, use node init.js and node server.js instead.

Check setup anytime: callwright-init status (or node init.js status). Place a call from the CLI (bypassing MCP): node dispatch.js <job.json> --go.

For local use, MCP_AUTH_TOKEN is optional — with it unset the server accepts loopback connections only. Set it once you expose the server beyond localhost (see Deploy hosted).

Deploy hosted (single-user)

Goal: a remote MCP URL only you can use. Works in both web clients (claude.ai, chatgpt.com) and local clients.

1. Pick a host with a persistent disk

config.json and scenario-profiles.json are written at runtime (by configure and learn_from_call), so avoid ephemeral/serverless filesystems. Good fits: Fly.io (volume), Railway (volume), Render (disk). A small always-on instance is fine — call volume is tiny.

2. Set environment variables (host dashboard)

RETELL_API_KEY = key_xxx            # your Retell secret - stays server-sideMCP_AUTH_TOKEN = <long random>      # the bearer token only you holdPORT           = 8787               # or whatever the host injectsCALLWRIGHT_DATA_DIR = /data         # persistent volume for runtime stateCALLWRIGHT_PUBLIC_URL = https://your-host.example.com   # optional: enables the auto-learn webhook (Retell posts call_analyzed back here)CLAUDE_ROUTINE_URL = https://api.anthropic.com/v1/claude_code/routines/trig_xxx/fire   # optional immediate call completion wake-upCLAUDE_ROUTINE_TOKEN = <routine trigger token>   # dedicated token generated by the Claude Routine

See .env.example. Never put the Retell key in a browser or chat.

When a call originates from a TickTick task, the calling assistant can pass tracking.ticktick_task_id to place_call. Callwright stores that opaque ID in Retell's storage-only metadata, preserves it through automatic retries, and uses the final call_analyzed webhook to wake the configured Claude Routine. The Routine URL and token are optional; hourly polling remains the fallback.

3. Deploy

Docker is provided:

docker build -t callwright .docker run -p 8787:8787 --env-file .env callwright

Or point the host at this repo (it auto-detects the Dockerfile). Mount a volume at the app dir so config.json / scenario-profiles.json persist across restarts.

No-Docker alternative: install from npm on the host and run the binary directly (point CALLWRIGHT_DATA_DIR at your persistent volume so runtime state survives restarts):

npm install -g callwrightCALLWRIGHT_DATA_DIR=/data callwright     # reads RETELL_API_KEY / MCP_AUTH_TOKEN from env

4. First-run setup over MCP (no webpage needed)

Do setup in chat:

1. Connect the server (below).
2. Ask the assistant to call configure with run_provisioning: true and your name + callback number. That verifies the key, selects your Retell number, creates the generic agent, and saves your profile.
3. get_setup_status should now report ready: true.

(Or run node init.js once on the host shell if you prefer.)

Connect a client

Replace the URL with your deployed host and the token with your MCP_AUTH_TOKEN. The server accepts the token two ways: an Authorization: Bearer header (Desktop/Code/Cursor) or a ?key= URL query param (web clients whose connector form has no header field).

Claude.ai (web)

The web connector form only takes a URL, so put the token in the URL:

1. Settings → Connectors → Add custom connector (or Browse connectors → Add custom).
2. Name: Callwright
3. URL: paste your host + /mcp + your token as ?key=:
   https://your-host.example.com/mcp?key=<MCP_AUTH_TOKEN>

4. Save, then enable it in a chat (the 🔌/tools menu). Ask "what callwright tools do you have?" to confirm it connected.

Because the token is in the URL, treat that connector URL like a password — anyone with it can place calls billed to your Retell account. Rotate by changing MCP_AUTH_TOKEN and updating the connector.

Claude Desktop / Cursor

{  "mcpServers": {    "callwright": {      "url": "https://your-host.example.com/mcp",      "headers": { "Authorization": "Bearer <MCP_AUTH_TOKEN>" }    }  }}

Claude Code

claude mcp add --transport http callwright https://your-host.example.com/mcp \  --header "Authorization: Bearer <MCP_AUTH_TOKEN>"

ChatGPT (Developer mode / connectors)

Add a custom MCP connector with the same URL — header or ?key= both work.

Keep "confirm before running tools" on. place_call is already dry-run-by-default, but tool-confirmation is a good second gate for an actuator that spends money and reaches real people.

Security notes

• Single bearer token = your lock. Anyone with the token can place calls billed to your Retell account. Treat it like a password; rotate by changing MCP_AUTH_TOKEN.
• Least-privilege Retell key. Use a scoped key if you don't need full admin.
• PII. Standing facts are stored in config.json on your host. Calls send only the minimal subset, shared only if the business asks; the read-back shows the footprint every time.
• If MCP_AUTH_TOKEN is unset, the server accepts loopback only (local dev).

Key files

Per-user runtime files — config.json, agents.json, .retell-ids.json, job.*.json, grounding.*.json — are git-ignored. They're created at runtime (by init/configure) or are personal call data; never commit them. Start from config.example.json and examples/.

A note on disclosure & legality

The agent always discloses that it's an AI at the start of every call — this is a hard, non-removable behavior. AI-call disclosure is legally required in some jurisdictions (e.g. California's B.O.T. Act) and is good practice everywhere. The agent also never fabricates information and shares personal data only when a business asks for it. For commercial or at-scale use, confirm requirements with counsel in your jurisdiction.

Optional: proactive email on every call

See webhook/ — a serverless Retell→Resend notifier that emails you each outcome. It is independent of the MCP (which retrieves outcomes on demand via get_call_outcome) and is not deployed or wired by default.