eodia/messagerie repository overview

⁠Messagerie

Open-source customer messaging, with an AI agent on the front line and a copilot for your agents. One container, one PostgreSQL: everything else is configured in the inbox.

This image runs both halves of Messagerie:

• the inbox for your agents, on port 3210: conversations, the copilot and all of the administration;
• the server, on port 8810: the API, sign-in, real time, the widget script (/widget.js), the REST API and the MCP server.

Alongside it, all it needs is PostgreSQL 16 with pgvector. No Redis, no message queue: real time and background jobs go through PostgreSQL itself.

📖 Documentation⁠ (in French) · 💻 Source code⁠ · ⚖️ License AGPL-3.0-or-later⁠

⁠What it does

• The widget, one script to paste into your website. The AI agent answers from your knowledge base and cites its sources; it calls your tools (an API, an MCP server, the visitor's record) and hands over to a human agent, with a summary, as soon as it is unsure.
• The inbox: inboxes, teams, transfers, tags, canned replies, a Ctrl+K command palette, attachments read by the AI on request, desktop and mobile notifications.
• The copilot drafts a reply for the agent; the agent's verdict (accepted, edited, rejected) feeds the evaluation set.
• Channels: the widget, e-mail (IMAP and SMTP), SMS and RCS (Twilio, SMS Mode).
• Automations: a trigger, conditions and steps — assign, tag, reply, notify, call a URL, ask the AI, wait.
• Integrations: REST API /api/v1, MCP server /mcp, signed webhooks, a JavaScript API for the widget, and page actions the AI can run with the visitor's consent.
• Dashboards: guided questions or SQL, over read-only analytics views.
• Accounts: password, invitation links, or your company's identity provider (OpenID Connect).
• AI: Mistral by default, or OpenAI, Ollama, any OpenAI-compatible server. Personal data is masked before it is sent to an external model. Without a key, conversations go straight to your agents.

⁠Try it in two minutes

All you need is Docker with Compose v2. In an empty folder, create this docker-compose.yml:

name: messagerie-trial

services:
  postgres:
    image: pgvector/pgvector:pg16
    environment:
      POSTGRES_USER: messagerie
      POSTGRES_PASSWORD: messagerie
      POSTGRES_DB: messagerie
    volumes:
      - postgres-data:/var/lib/postgresql/data
    healthcheck:
      test: ["CMD-SHELL", "pg_isready -h 127.0.0.1 -U messagerie -d messagerie"]
      interval: 5s
      timeout: 3s
      retries: 30

  messagerie:
    image: eodia/messagerie:${MESSAGERIE_VERSION:-latest}
    depends_on:
      postgres: { condition: service_healthy }
    environment:
      DATABASE_URL: postgres://messagerie:messagerie@postgres:5432/messagerie
      CHAT_SECRET: ${CHAT_SECRET:?set CHAT_SECRET in .env}
      CHAT_WEB_ORIGIN: http://localhost:3210
      CHAT_PUBLIC_URL: http://localhost:8810
      CHAT_API_URL: http://localhost:8810
      # Optional: without a key, conversations go to your agents.
      CHAT_AI_API_KEY: ${CHAT_AI_API_KEY:-}
    ports:
      - "127.0.0.1:3210:3210"
      - "127.0.0.1:8810:8810"
    volumes:
      - files:/data/files

volumes:
  postgres-data:
  files:

Then:

echo "CHAT_SECRET=$(openssl rand -base64 32)" > .env
docker compose up -d

Open http://localhost:3210⁠: the welcome screen creates the first supervisor. Then set up a site, an inbox and a team under Administration (see Premiers pas⁠).

To let the AI answer, add a Mistral key and restart:

echo "CHAT_AI_API_KEY=…" >> .env
docker compose up -d

⁠With your own PostgreSQL

docker run -d --name messagerie \
  -p 3210:3210 -p 8810:8810 \
  -v messagerie-files:/data/files \
  -e DATABASE_URL=postgres://messagerie:…@db.example.com:5432/messagerie \
  -e CHAT_SECRET="$(openssl rand -base64 32)" \
  -e CHAT_WEB_ORIGIN=http://localhost:3210 \
  -e CHAT_PUBLIC_URL=http://localhost:8810 \
  -e CHAT_API_URL=http://localhost:8810 \
  eodia/messagerie

The DATABASE_URL user must be able to create the vector extension (or find it already created). For SQL questions in dashboards, it must also be able to create the chat_analytics role; without it, only guided questions work.

⁠The image

On startup, the server applies the migrations of the chat schema: on an empty database it creates it; afterwards, it only adds what is missing. If either the server or the inbox stops, the whole container stops, and the restart policy brings it back.

⁠Tags

A prerelease (0.2.0-rc.1) is published under its exact tag only. In production, pin an exact version rather than following latest.

⁠Environment variables

⁠Required in production

⁠Optional

The full list, with defaults: Variables d’environnement⁠.

No secrets in the configuration. An AI tool, an MCP server, an SMS number or an e-mail address names the variable that holds its secret (${WEATHER_TOKEN}), never the value itself: define that variable in the container's environment.

⁠In production

The same image, behind an HTTPS gateway. The inbox and the server each get their own name, two subdomains of the same domain (support.example.com and chat.example.com): the agent's session is a cookie set by the server, which the browser only sends within the same site.

The Mise en production⁠ guide gives the complete docker-compose.yml — PostgreSQL, Messagerie, the optional worker and Caddy for certificates — along with backups and upgrades.

To upgrade:

docker compose pull && docker compose up -d   # migrations are applied on startup

⁠License

Open-source software by Eodia⁠, released under the GNU Affero General Public License v3.0⁠ or any later version.