Agnt is a minimalist, autonomous AI agent CLI tool written in the Odin programming language. It follows a modular, hexagonal architecture designed for extreme extensibility, privacy, and efficiency.
- Hexagonal Architecture: Ports and Adapters for LLM providers and communication channels.
- Unified Provider: Connects to ANY OpenAI-compatible API (OpenRouter, Groq, LM Studio, Ollama, Z-AI).
- Multi-Channel: Interact via terminal (CLI) or Telegram Bot.
- Interactive Setup: Simple onboarding with
--setupcommand and model discovery. - Hybrid Memory System:
- Static: Loads context from
.mdfiles in your workspace (IDENTITY.md,SOUL.md). - Learned: Persistent facts stored in a local SQLite database.
- Graph (Barabási-style): Semantic relations between entities (subject-predicate-object) mapped in SQLite.
- Static: Loads context from
- Agent Tools (Hands): Autonomous file I/O and real-time web search/fetch capabilities.
- Odin Compiler
- libsqlite3-dev (
sudo apt install libsqlite3-dev) - curl
git clone https://github.com/manelsen/agnt
cd agnt
# Build normal
odin build src -out:agnt
# Build otimizado (apenas ~450KB)
odin build src -out:agnt -o:speed
strip agnt
./agnt --setupAgnt can run as a background service on Linux:
agnt service install: Installs the systemd user service.agnt service uninstall: Removes the service.agnt service start / stop / restart: Controls the service.agnt service status: Checks if the agent is running.
Agnt keeps your system clean by isolating its brain in:
- Config:
~/.agnt/config/.env - Memory:
~/.agnt/workspace/
- Multi-Agent Protocol: Allow Agnt to delegate tasks to other sub-agents and coordinate complex workflows.
- Shell & Python Tools: Autonomous shell command execution and a sandboxed Python interpreter for data processing.
- Semantic Vector Memory: Integration with
sqlite-vecfor RAG. - Voice Interface: Audio-to-text and text-to-speech adapters.
- Graph Visualization: Export mental maps to Obsidian/Logseq.
MIT