EvoNexus on Docker Swarm

This repo supports three independent deployment paths. Each one uses a different set of files, and none of them interferes with the others:

Path	Files used	Who it's for
VPS bare-metal — git clone + make setup + install-service.sh	setup.py, install-service.sh, Makefile	Single-VPS installs with systemd
Local dev with Docker Compose — docker compose up	Dockerfile, Dockerfile.dashboard, docker-compose.yml	Local development, single-machine demos
Production on Docker Swarm — Portainer / Traefik	Dockerfile.swarm, Dockerfile.swarm.dashboard, evonexus.stack.yml	Multi-node production behind a reverse proxy

This guide covers the third path. The first two are documented elsewhere and are not modified by the Swarm overlay.

Architecture

Three Swarm services in one stack:

Service	Image	Purpose
evonexus_dashboard	evo-nexus-dashboard	Flask + React + embedded terminal-server, exposed through Traefik
evonexus_telegram	evo-nexus-runtime	Long-running claude --channels plugin:telegram… daemon
evonexus_scheduler	evo-nexus-runtime	Background scheduler for ADW routines (scheduler.py)

Seven named volumes (auto-created by Portainer on first deploy):

• evonexus_config — the UI-editable .env and providers config (writable)
• evonexus_workspace — agent workspace (daily-logs, projects, finance, etc.)
• evonexus_dashboard_data — dashboard SQLite
• evonexus_memory — persistent agent memory
• evonexus_adw_logs — routine execution logs
• evonexus_agent_memory — per-agent scratch memory
• evonexus_codex_auth — ~/.codex/auth.json persisted across restarts

One external overlay network (traefik-public by default — rename to match your Traefik setup).

UI-first configuration

The stack intentionally ships zero API keys and zero integration tokens. Everything is configured through the dashboard after the first boot:

• Providers (/providers) — Anthropic, OpenAI API Key, Codex OAuth, OpenRouter, Gemini, etc.
• Integrations (/integrations) — Telegram, Stripe, Omie, Bling, Asaas, Fathom, Todoist, GitHub, Linear, Gmail, Calendar, YouTube, Instagram, LinkedIn, Evolution API, Evo CRM, …
• Env editor — anything else

When you save a value in the dashboard, it lands in /workspace/config/.env inside the evonexus_config volume. On the next container start, the entrypoint sources that .env so every process sees the value as a regular environment variable.

Runtime daemon services — what does and doesn't need its own container

Only channels that maintain a persistent connection need a dedicated long-running service. External REST APIs do not.

Channel	Dedicated service?	Notes
Telegram	✅ Included (evonexus_telegram)	Already in the stack
Discord	Optional	Duplicate the evonexus_telegram block; swap plugin:telegram → plugin:discord
iMessage	Optional	Same pattern; needs a BlueBubbles server on a Mac
WhatsApp (Evolution API / Go)	❌ Not needed	External HTTP API — run the evolution_api stack separately
Stripe / Omie / Bling / Asaas / Fathom / Todoist	❌ Not needed	External REST APIs, called on demand
GitHub / Linear / Gmail / Calendar / YouTube / Instagram / LinkedIn	❌ Not needed	OAuth via dashboard, REST calls on demand
Evo CRM	❌ Not needed	Reached via cross-stack HTTP over the shared Traefik network

Deploy flow

Prerequisites

• Docker Swarm initialized on at least one manager node.
• A Traefik stack already running, attached to an external overlay network (this guide assumes traefik-public — rename to match your setup).
• Traefik configured with a websecure entrypoint and a certificate resolver called letsencryptresolver (rename in the stack labels if your resolver has a different name).
• A DNS A record pointing evonexus.<yourdomain> to the Swarm ingress.
• Docker images are pulled from Docker Hub at evoapicloud/evo-nexus-{dashboard,runtime} (published by the included GitHub Actions workflow on every version tag and push to main).
• The images are public, so no docker login is needed on the Swarm managers. If you fork the project and publish to your own namespace, update the image: lines in evonexus.stack.yml.

Steps

1. Build and publish the images. Push a version tag (vX.Y.Z) or use the Actions tab to run Build & Publish Docker Images (Swarm) manually. The workflow builds both images in parallel and publishes them with the version tag and :latest.

2. Open Portainer → Stacks → Add stack.

   • Name: evonexus
   • Paste the contents of evonexus.stack.yml
   • Replace the placeholders: evonexus.example.com (your hostname, two places) and traefik-public (your Traefik overlay network name, several places). The images already point at the official evoapicloud/evo-nexus-* namespace on Docker Hub.
   • Click Deploy

3. Watch the containers come up. evonexus_dashboard serves the SPA immediately. evonexus_telegram and evonexus_scheduler will log waiting for ANTHROPIC_API_KEY every 30 seconds — this is expected until step 4.

4. Open https://evonexus.<yourdomain> → Setup wizard → create the admin user → Providers → pick and save a provider:

   • Anthropic — paste sk-ant-… (default, no extra setup).
   • OpenAI API Key — paste sk-… (create at platform.openai.com/api-keys with All permissions, not Restricted).
   • Codex OAuth — click the Login button and complete the ChatGPT device flow. The token is written to the evonexus_codex_auth volume and persists across restarts.

5. Wait ~30 seconds. The telegram and scheduler services detect the key on their next polling cycle and start running.

6. Integrations. Configure channels and external APIs from /integrations. Values land in the same config volume and are picked up at container startup. After changing something the telegram daemon uses, force-restart that service in Portainer (docker service update --force evonexus_telegram).

Troubleshooting

Terminal says "Could not reach terminal-server"

Confirm the dashboard container logs show both processes on boot:

[start-dashboard] terminal-server on :32352, Flask on :8080
🚀 Terminal server running at http://localhost:32352
 * Running on http://127.0.0.1:8080

If only Flask is running, the image is an old build — redeploy with the latest tag. Also verify that the Traefik router includes the evonexus_terminal_strip middleware (visible in the Traefik dashboard if you have one enabled).

telegram / scheduler crash-looping

They should not crash — they wait for ANTHROPIC_API_KEY via the entrypoint's poll loop. If they exit immediately, the image predates the bootstrap entrypoint. Redeploy with a current tag.

Volumes lost after stack redeploy

Portainer preserves named volumes across stack updates by default. If you removed and recreated the stack, data in evonexus_config, evonexus_workspace, etc. is gone unless you backed them up. Confirm with docker volume ls on the manager.

What is NOT changed from the main codebase

The Swarm overlay is purely additive. Every upstream file is unmodified and continues to support the VPS and local Compose paths:

• Dockerfile, Dockerfile.dashboard, docker-compose.yml
• setup.py, install-service.sh, Makefile
• pyproject.toml, uv.lock
• All Python, React, skills, agents, ADWs, configs

Swarm-specific files added by this overlay:

• Dockerfile.swarm, Dockerfile.swarm.dashboard
• entrypoint.sh, start-dashboard.sh
• evonexus.stack.yml
• .github/workflows/docker-publish.yml
• README.swarm.md (this file)