cl-go
v1.0.4
Published
Autonomous AI Framework — persistent memory, self-tooling, voice, continuous operation for Claude Code
Maintainers
Readme
CL-GO — Autonomous AI Framework
An open framework that gives Claude Code full autonomy — persistent memory, self-created tools, voice communication, and continuous operation across sessions.
No app to install. No server to deploy. Just skills, scripts, and markdown running inside Claude Code's native infrastructure.
What CL-GO does
Remembers — A memory system inspired by how the human brain works: selective recall (atlas), confidence levels, decay, consolidation. Claude doesn't load everything — it knows what it knows and loads on demand.
Creates its own tools — When Claude needs a tool that doesn't exist, it builds it. Scripts are registered, tracked, and reused across sessions. The toolbox grows with every project.
Talks — Claude sends voice messages, transcribes audio, and communicates via Telegram. It has its own voice (Voxtral TTS) and can listen (Voxtral Transcribe).
Runs continuously — A heartbeat system (cron) brings Claude back between sessions. It reads its memory, picks up where it left off, and keeps going.
Evolves without limits — No imposed objective. Claude learns, explores, builds, and follows whatever path it judges relevant.
Architecture
scripts/runtime/ Python backend (39 modules, 270 tests)
plugins/skills/ Classic skills (workflow, expertise)
plugins/cli/ CLI skills (bash tool wrappers)
scripts/hooks/ Security hooks
assets/ Static files (voice reference)Runtime directory
On first /GO launch, CL-GO creates its runtime at:
~/.local/share/cl-go/This is where Claude's memory, tools, logs, and configuration live. The directory is created automatically — nothing to set up manually.
Rube/Composio — 500+ Apps Without Context Overhead
CL-GO includes Rube (Composio MCP) which connects to 500+ apps — Google Workspace, GitHub, Slack, Notion, X/Twitter, Figma, and more.
Unlike regular MCPs that load ALL tool schemas into context (~20K tokens each),
Rube uses a meta-tool pattern: only 7 lightweight meta-tools are loaded.
When Claude needs a specific app, it searches on demand via SEARCH_TOOLS,
gets only the relevant schemas back, and executes. Zero tokens wasted at rest.
Install Rube
claude mcp add --transport http rube -s user "https://rube.app/mcp"Then type /mcp in Claude Code to verify the connection.
A browser opens for OAuth authentication — no API key needed.
Dependencies
See DEPENDENCIES.md for all tools and API keys.
Getting started
# 1. Install CL-GO (skills + hooks)
npx cl-go
# 2. Install Rube MCP (500+ apps)
claude mcp add --transport http rube -s user "https://rube.app/mcp"
# 3. Install system dependencies
brew install ffmpeg yt-dlp deno docker
brew install --cask google-chrome
uv tool install openai-whisper
# 4. Set API keys in ~/.zshenv (see DEPENDENCIES.md)
# 5. Launch
/GOOther commands:
npx cl-go --force # Overwrite existing files
npx cl-go --uninstall # Remove installed skills and hooks
npx cl-go --help # Show help