Architecture

Astervis is a small stack of Docker containers that runs on the same server as your PBX. It taps into the call records Asterisk already writes, copies them into its own analytics database, and serves a dashboard — without ever touching the cloud. This page is the mental model: what the pieces are, how data moves between them, and where the AI fits in.

The data flow

A call finishes, Asterisk writes a CDR row into its MySQL/MariaDB database, and from there Astervis takes over. The CDC service notices the new row in the binary log, hands it to the replication consumer, and it lands in TimescaleDB. The backend and admin read from TimescaleDB and render the dashboard you open on port 8443.

If the call has a recording and AI is enabled, a second path kicks in: the queue picks up the audio, sends it to the AI gateway (Gemini on Vertex AI), and the resulting transcript and analysis are written back into the same database — so they show up on the call card alongside the metrics.

  ┌─────────────────────┐
  │  Asterisk / FreePBX │   call ends → CDR row + (optional) recording
  │   MySQL CDR  (host) │
  └──────────┬──────────┘
             │ binary log
             ▼
  ┌─────────────────────┐
  │  litcdc  (CDC)      │   reads binlog, snapshots then streams
  └──────────┬──────────┘
             │ HTTP long-poll feed
             ▼
  ┌─────────────────────┐
  │  replication        │   consumes feed → writes rows
  └──────────┬──────────┘
             ▼
  ┌─────────────────────┐        ┌──────────────────────────────┐
  │  TimescaleDB        │◀──────▶│  backend (app) + admin (web) │
  │  (Postgres + TS +   │        └───────────────┬──────────────┘
  │   pgvector)         │                        │
  └──────────┬──────────┘                        ▼
             │                          ┌──────────────────┐
             │  recordings              │  nginx  :8443    │  ← you, the dashboard
             ▼                          └──────────────────┘
  ┌─────────────────────┐   audio    ┌──────────────────────────┐
  │  queue (BullMQ)     │───────────▶│  AI gateway (Gemini)     │
  └─────────────────────┘            └────────────┬─────────────┘
             ▲                                     │ transcript + analysis
             └─────────────────────────────────────┘  (back into TimescaleDB)

What gets installed

The installer renders a docker-compose.yml into /opt/astervis and brings up the following services. Only nginx is reachable from outside the machine; everything else binds to 127.0.0.1.

On disk it looks like this:

litcdc — the CDC pipeline

CDC (change data capture) is how Astervis keeps its analytics database in lockstep with Asterisk without polling MySQL and without slowing the PBX down. The whole job is done by litcdc, a single Rust binary (~10–25 MiB resident) that reads the MySQL/MariaDB binary log directly.

It runs in two phases:

The replication consumer pulls those events over an HTTP long-poll feed (litcdc listens on litcdc:8088 inside the Docker network, published to 127.0.0.1:18088 on the host) and writes them into TimescaleDB. The buffer is capped at 64 MiB; once events are consumed past that cap they're truncated, and the writer pauses if nothing is draining — so a stalled consumer can never blow up disk usage.

litcdc replaced the old Debezium + Redpanda pipeline. That pair needed a JVM and a full message broker and cost roughly 640 MiB of RAM; litcdc does the same job — durable buffering, offsets, replay, backpressure — in around 30 MiB, with no broker and no JVM. That single change is why the whole stack now fits in ~0.6 GB and the RAM minimum dropped to 2 GB.

The AI pipeline (optional)

Analytics work fully on their own — the AI pipeline is a separate, optional layer that turns recordings into transcripts, QA scores, and searchable text. It only runs if you hold an AI license tier, and it's the only part of Astervis that sends data off the server.

When a CDR with a recording lands, the queue worker submits the audio to the AI gateway, which runs a single multimodal Gemini call on Google Vertex AI — transcription, speaker diarization, sentiment, category, and (if configured) rubric-based QA all in one request. Results come back via webhook into the ai_transcripts table, moving through pending → processing → completed. From there they're embedded with gemini-embedding-001 into pgvector for semantic search.

Nothing about AI is required. If your calls can't leave the building, simply don't enable it — the dashboards, CDR replication, and integrations all work without it.

Resource footprint

The full stack sits at roughly 0.6 GB of RAM at rest — most of the savings come from litcdc replacing the old broker-based CDC. The hard minimum to install is 2 GB of RAM, 1 CPU core, and 15 GB of free disk (checked on the Docker data partition, not /). 4 GB and 2 cores are recommended for comfortable headroom during the initial snapshot.

Network surface

This is the part worth internalizing for anyone hardening the box: Astervis exposes a single port.

Where to go next