CLIO (Command Line Intelligence Orchestrator) is a terminal coding agent written in pure Perl with only standard core modules. It works as a pair programmer: you describe a task, it investigates the codebase, proposes a plan, and after your approval edits files, runs tests, and commits. You can interrupt it mid-task with any key, and it pauses to ask what you need.
Because it needs only Perl and standard Unix tools, CLIO runs in local shells, SSH sessions, tmux, Docker, and headless servers. It can deploy itself to remote machines over SSH and run a task there, and it can spawn parallel sub-agents coordinated through file locks, git locks, and rate limiting, with live output panes in tmux, GNU Screen, or Zellij.
Sessions persist with full history, long-term memory carries across projects, and a profile learns the user's working style. Skills load from Git repositories, plugins add tools, MCP servers connect external tools, and OpenSpec supports spec-driven development. Every turn's changes can be undone.
CLIO ships 16 provider configurations. All API providers are OpenAI-compatible, and Anthropic, Google, and NVIDIA also have native protocol adapters. Providers include GitHub Copilot (OAuth), Anthropic, OpenAI, Google Gemini, DeepSeek, OpenRouter, Ollama Cloud, MiniMax, Z.AI, NVIDIA NIM, Vercel AI Gateway, and the local servers llama.cpp and LM Studio.
Security controls include secret redaction before context reaches the model, risk classification of shell commands with approval for high-risk ones, permission prompts for file access outside the project, a sandbox mode, Docker container isolation, and filtering of invisible Unicode characters used for prompt injection.
Features
- Plan-then-implement workflow: investigates, proposes a plan, and implements only after approval, with interruption at any point
- Git operations: status, diff, commit, branch, push, pull, stash, tag, and worktree
- Multi-agent coordination: parallel sub-agents with file locks, git locks, rate limiting, and an inbox
- Multiplexer panes: live agent output in tmux, GNU Screen, or Zellij
- Remote execution: runs AI tasks on other machines over SSH
- Memory and profile: persistent sessions, long-term memory across projects, and a learned working-style profile
- Undo: reverts AI changes from any turn
- Extensibility: skills from Git repositories, drop-in plugins, MCP servers, and OpenSpec spec-driven development
- 16 providers: GitHub Copilot, Anthropic, OpenAI, Gemini, DeepSeek, OpenRouter, MiniMax, Z.AI, NVIDIA NIM, llama.cpp, LM Studio, and others
- Security layers: secret redaction, command risk classification, path authorization, sandbox mode, container isolation, and invisible-character filtering
- No package dependencies: pure Perl with standard core modules
