I wanted an AI agent that actually grows with you. Not a chatbot. Not a wrapper around tool-calling. Something that runs on your machine, has real access to your system, and gets more capable over time by building its own tools.
So I started building EloPhanto. It starts with 78+ tools (filesystem, shell, 47 browser automation tools using your actual Chrome profile, document analysis with OCR, scheduling). When you ask it to do something it can't, it enters a self-development pipeline: researches the problem, designs a solution, implements it, writes tests, reviews its own code with a different model, and deploys — all autonomously.
It remembers everything across sessions. It has an evolving identity that develops through reflection on completed tasks. It can decompose complex goals into checkpoints and work on them over days. It's also aware of its own code and files - be careful with it!
You can talk to it from your terminal, Telegram, Discord, or Slack — all through a WebSocket gateway with isolated sessions. There's a skills system with 27 bundled best-practice guides and a pending public registry (EloPhantoHub) for community sharing.
Security: encrypted credential vault, three-tier permissions (ask-always → smart-auto → full-auto), and more.
Multi-model: routes tasks across OpenRouter (Claude, GPT), Ollama (local), and Z.ai/GLM - coding plan - based on task type and cost.
Next up: reached out to the Unsloth team on a self-learning pipeline so EloPhanto can fine-tune from its own task history.
petrroyce•1h ago
So I started building EloPhanto. It starts with 78+ tools (filesystem, shell, 47 browser automation tools using your actual Chrome profile, document analysis with OCR, scheduling). When you ask it to do something it can't, it enters a self-development pipeline: researches the problem, designs a solution, implements it, writes tests, reviews its own code with a different model, and deploys — all autonomously.
It remembers everything across sessions. It has an evolving identity that develops through reflection on completed tasks. It can decompose complex goals into checkpoints and work on them over days. It's also aware of its own code and files - be careful with it!
You can talk to it from your terminal, Telegram, Discord, or Slack — all through a WebSocket gateway with isolated sessions. There's a skills system with 27 bundled best-practice guides and a pending public registry (EloPhantoHub) for community sharing.
Security: encrypted credential vault, three-tier permissions (ask-always → smart-auto → full-auto), and more.
Multi-model: routes tasks across OpenRouter (Claude, GPT), Ollama (local), and Z.ai/GLM - coding plan - based on task type and cost.
Next up: reached out to the Unsloth team on a self-learning pipeline so EloPhanto can fine-tune from its own task history.
Python core, TypeScript browser bridge. Apache 2.0. One-command setup.
It's still in development so excuse any issues. Enjoy!
https://github.com/elophanto/EloPhanto