Connect by JBRH Open Connect

The Connect technology stack

Connect is a Python application. FastAPI runs under Uvicorn, SQLAlchemy reaches PostgreSQL, and a no-build browser frontend talks to it. It is hosted on an AWS server behind Caddy, and systemd supervises it. LiveKit and Connect AI's realtime speech model carry live calls over a Vobiz phone line. Mail goes through Stalwart, which relays outbound mail through SMTP2GO, WhatsApp through Meta's Cloud API, and payments through Cashfree. This page is the inventory; each row links to a page of its own.

Status
Reference What this means
Audience
both, developer
Last verified
Product version
6.3.2

How to read this library#

The rest of the technology reference explains ideas and standards: SIP, DKIM, row-level security, retrieval. This group is different. It covers the products and libraries Connect is built from, found by reading the repository's own dependency files, vendored code and host scripts rather than from a list of what such a product usually uses.

Every page has the same order, so the same answer is always in the same place:

  1. What it is. For anyone, in plain words, with the category and the official site.
  2. Where Connect uses it and how it connects. The files and the path a request takes.
  3. Benefits and limits, security and privacy, and licence, cost and versions.
  4. Troubleshooting and the official references.

The request path, end to end#

A request from a browser passes through six steps to reach the database. Each one is a page in this library.

  1. Browser: plain ES modules, no build step (the browser frontend)
  2. Caddy: TLS, compression and the reverse proxy
  3. Uvicorn: the ASGI server on the loopback port
  4. FastAPI and Starlette: routing, middleware and validation through Pydantic
  5. SQLAlchemy and psycopg: the ORM, the workspace kernel and the driver
  6. PostgreSQL 16 on the same host, with row-level security

A phone call takes a second path that does not pass through the web server at all:

  1. The caller dials a number on the Vobiz carrier
  2. The carrier's SIP trunk delivers the call into a LiveKit room
  3. The connect-voice worker (systemd) joins the room
  4. A realtime speech model session listens and speaks
  5. At the end the worker posts the call to the application, which writes it to PostgreSQL

Where Connect uses each piece#

TechnologyCategoryIn Connect (as recorded)Licence
PythonLanguage3.12 on the host (ops/aws/bootstrap-host.sh)PSF License
FastAPI and StarletteWeb frameworkfastapi>=0.115MIT; BSD-3-Clause
UvicornASGI serveruvicorn[standard]>=0.30BSD-3-Clause
SQLAlchemy and psycopgORM and database driversqlalchemy>=2.0, psycopg[binary]>=3.2MIT; LGPL-3.0
PostgreSQLDatabase16 (host packages)PostgreSQL License
PydanticData validationpydantic>=2.7MIT
HTTPXHTTP clienthttpx>=0.27BSD-3-Clause
cryptography and argon2-cffiCryptographycryptography>=42, argon2-cffi>=25.1,<26Apache-2.0 or BSD-3-Clause; MIT
Locale librariesPhone numbers, holidays, time zonesphonenumbers>=9.0.41, holidays>=0.106, tzdata>=2024.1Apache-2.0; MIT; Apache-2.0
LiveKit AgentsRealtime medialivekit-agents==1.8.0Apache-2.0
Realtime speech modelModel API (Connect AI)through the LiveKit provider plugin (1.8.0) and HTTPSCommercial service
VobizTelephony carrieradapter in voice_carriers.pyCommercial service
WhatsApp Cloud APIMessaging APIGraph API v21.0Commercial service
Cashfree PaymentsPayment gatewayAPI version 2025-01-01Commercial service
Amazon SESEmail sending, where an identity is configured for itSES v2 through boto3>=1.35Commercial service; boto3 Apache-2.0
StalwartMail server; relays outbound mail through SMTP2GOon the host (ops/mail/)AGPL-3.0 or the Stalwart Enterprise License
AWS hostingCloudEC2, EBS, S3, SSM in ap-south-1Commercial service
CaddyWeb serverops/CaddyfileApache-2.0
systemdService managerconnect, connect-voiceLGPL-2.1-or-later
PlaywrightBrowser automationplaywright==1.63.0 (research, tests)Apache-2.0
The browser frontendFrontendES modules, no bundlerConnect's own code
JsSIPSIP over WebSocket3.13.8, vendoredMIT
three.js3D graphicsr169 (0.169.0), vendoredMIT
GSAP, Lenis and AutoAnimateMotionGSAP 3.15.0, Lenis 1.3.26, vendoredGSAP Standard License; MIT; MIT
Chrome Manifest V3Browser extensionMarketplace Assistant 1.1.0Connect's own code
Development toolsQuality gatesRuff, Bandit, the regression runnerMIT; Apache-2.0

Some adapters exist in the code for providers a business may choose: Twilio, Plivo and Exotel for calls, MSG91 for SMS, Razorpay and others for payments, Gmail and Microsoft Graph for connected mailboxes. They are covered where the capability is documented. This inventory is the platform's own stack.

What is deliberately absent#

  • No JavaScript build step. There is no package.json and no bundler. The browser loads the files it is served, and third-party browser code is vendored with a recorded checksum.
  • No container orchestrator. One server runs two systemd services, PostgreSQL, Caddy and the mail server. A release is an atomic directory swap with a verified database dump before it.
  • No second database. PostgreSQL holds every record. SQLite appears only as a development convenience and is not a production path.
  • No secrets on these pages. Credentials live in the host's environment file or a secret store, and each page names the setting, never the value.

Questions#

How was this inventory produced?

From the repository: the three requirements files, the vendored-code record, the host bootstrap and service files, the Caddyfile, and the imports that the backend actually makes. Babel and pycountry, for example, are imported only by an offline data builder, so they appear on the locale page with that caveat rather than as runtime dependencies.

Does Connect use a large language model from one provider only?

No. Connect AI is the product name, and the model underneath each feature is a setting the platform operator chooses, with adapters for several providers. The realtime speech model page covers the live voice. Choosing a model explains how a default is set.

Why are some versions exact and others floors?

The web application's dependencies are floors, so security releases arrive without a code change. The voice worker's packages are pinned exactly, because they were proven together on a real call, and the calling quality is locked: an upgrade there needs a measured call first.