vicode 0.4.0

TUI coding agent for local LLMs with supervisor system

vicode

TUI coding agent for local LLMs. Like Claude Code but runs on your own hardware.

Install

pip install vicode

Quick Start

# Start llama-server with your model
llama-server -m model.gguf -p 8080 --jinja

# Launch vicode in any project directory
cd my-project
vicode

On first run, vicode auto-detects your local LLM server. If none is found, an interactive setup wizard configures your provider (OpenAI, Anthropic, OpenRouter, Groq, or custom endpoint).

Features

• 13 tools: bash, file read/write/edit, grep, glob, ls, web search/fetch, memory, undo, explore, plan
• Supervisor system: Safety guardian validates actions against plans, memory guardian auto-stores learnings, task tracker shows progress
• Persistent memory: Remembers preferences, error fixes, and patterns across sessions (powered by engram)
• Persistent session traces: Stores structured traces for every TUI, CLI, and one-shot prompt session so you can inspect and improve agent behavior later
• Memory consolidation: Learns only from successful sessions, prunes noisy/redundant memories, and can be run as a daily maintenance task
• Sub-agents: Explore and Plan run with isolated context to avoid polluting the main conversation
• Slash commands: /help, /clear, /compact, /init, /undo, /model, /provider, /status, /memory, /sessions, /resume, /trace, /learn, /consolidate
• Custom commands: Drop .md files in .vicode/commands/ for project-specific shortcuts
• Provider support: Local (llama-server, Ollama, LM Studio), OpenAI, Anthropic, OpenRouter, Groq

Configuration

• VICODE.md / CLAUDE.md / AGENT.md in project root are loaded into the system prompt
• ~/.vicode/config.json stores provider settings
• vicode --setup to reconfigure provider
• vicode --consolidate-memory --cwd <project> runs daily-style memory cleanup + learning consolidation for one project namespace

Requirements

• Python 3.11+
• A running LLM server (llama-server, Ollama) or API key for a cloud provider