GoGogot — Open-Source Self-Hosted AI Agent Written in Go
GoGogot is a lightweight, self-hosted personal AI agent written in Go. Deploy on your own server as a single ~15 MB binary — it runs shell commands, edits files, browses the web, manages persistent memory, and schedules tasks. Powered by Claude or MiniMax via Telegram. A lightweight alternative to OpenClaw in ~4,500 lines of Go.
Features
- Shell Access — Execute bash commands, manage files, run scripts on your server.
- Web Tools — Search with Brave, fetch pages, make HTTP requests, download files.
- Persistent Memory — Markdown-based memory the AI agent reads and writes across sessions.
- Task Scheduler — Cron-based self-scheduling, persisted across restarts.
- Single Binary — ~15 MB compiled Go binary. Only 7 dependencies. No runtime required.
- Telegram Bot — Multi-chat support, attachments, typing indicators. Extensible transport layer.
- Auto Compaction — Automatic context window compression when approaching token limits.
- Multi-Model — Claude and MiniMax out of the box. Switch with one environment variable.
Architecture
No frameworks, no plugins, no magic. One agentic loop in Go with good tools and a smart prompt. The LLM decides everything — when to plan, when to act, when to self-correct. The entire agent idles at ~10 MB RAM on any Linux server or cheap VPS.
GoGogot vs OpenClaw vs Nanobot
GoGogot compiles to a ~15 MB binary with 7 dependencies and ~10 MB RAM idle. OpenClaw requires ~1 GB install size, 800+ npm packages, and ~450 MB RAM idle. Nanobot needs ~150 MB with 30+ pip packages. GoGogot is the lightest self-hosted AI agent available — deploy it on a $4/month VPS with docker compose up.
Quick Start
Clone the repo, configure .env with your Telegram bot token and API keys (Anthropic for Claude, OpenRouter for MiniMax), and run docker compose up. Your personal AI agent is ready in 60 seconds.
View on GitHub — Open source, MIT license.