EliaAgent AI Agent for macOS

amy 17/09/2026

EliaAgent transforms OpenCode into a self-managing fleet of autonomous AI subworkers. Instead of manually prompting a single assistant, you deploy a Python FastAPI server in Docker that runs scheduled, monitored, and self-healing AI agents 24/7.

This free open-source solution isolates memory leaks, handles provider outages with exponential backoff, and offers effectively unlimited tokens via free OpenCode Zen models and rotating residential proxies.

Features

  • Autonomous Scheduling: Agents run on cron expressions (e.g., hourly weekdays) without human intervention.
  • Self-Healing Architecture: Automatically detects transient errors, waits, and resumes tasks from the last known state.
  • Resource Efficient: Runs in Docker with ~4 GB RAM, containing OpenCode’s memory usage and preventing system blowups.
  • Unlimited Tokens: Leverages free local models and proxy rotation to bypass API rate limits and costs.
  • Real-Time Monitoring: REST API and WebSocket streams provide live logs, reasoning traces, and tool call visibility.
  • Cross-Platform Clients: Control the fleet via EliaTopBar (macOS menu bar) or EliaAndroidApp.
  • Remote Access: Built-in Cloudflare Tunnel wizard enables secure remote management with shared auth tokens.

Quick Start

Clone & Install:

git clone https://github.com/vakandi/EliaAgent.git
cd EliaAgent/subworkers
./install_subworkers.sh

Launch Fleet:

./start_subworkers.sh

Verify Health:

curl http://localhost:5656/health

Requirements

  • OS: macOS (with Colima) or Linux (Docker).
  • RAM: 4 GB allocated to Docker.
  • Engine: OpenCode (free Zen models).
  • Tools: tmux for session management.

Adding Subworkers

Expand your workforce by adding three configuration files per agent. No new code is required, just define the personality, workspace, and schedule. Ideal for developers seeking privacy-focused, local-first AI automation for reporting, coding, and data processing.

Run lifecycle

  1. APScheduler fires → runner checks OpenCode health
  2. Session created in the agent’s isolated workspace (x-opencode-directory header)
  3. PROMPT.md sent as the user message → agent works (streamed live over WS)
  4. Completion = session idle and <promise>DONE</promise> marker in output
  5. Transient provider errors → random wait → “continue the tasks” re-injected into the same session (up to 3×)
  6. Failure → exponential backoff retries; alerts via macOS beep / ntfy.sh

License

MIT License

Resources & Downloads

《EliaAgent AI Agent for macOS》

Related articles