An autonomous, local-first AI coding, research, and security agent for your terminal.
OCLI runs a model of your choice — a local GGUF/MLX model or a cloud API — inside a real agent loop with real tools: shell, files, git, web search, and a shared terminal that both you and the AI type into. It ships as a single C++20 binary and presents a live split-pane TUI plus an optional LAN web dashboard.
It is tuned as an ethical red-team / security assistant as well as a general coding agent.
- Split-pane live view — conversation on the left, a shared PTY terminal on the right that both you and the AI drive.
- Many backends — local (
llama.cpp,MLX,Ollama), cloud (OpenAI,Anthropic,NVIDIA NIM), or a remote MPC server you connect to. - Qwythos-9B (1M context) as the default local model, with a one-command installer.
- Context-window progress bar and effort levels that scale how long the model thinks before it acts; automatic history compaction.
- LAN web dashboard, parallel worker sub-agents, and an autonomous
/loopmode. - Portable sharing — run the model server in Docker and get a public Cloudflare-tunnel URL any OCLI can connect to.
- macOS or Linux
- A C++20 compiler, CMake, and libcurl (+ pthreads)
- macOS:
xcode-select --installthenbrew install cmake curl - Debian/Ubuntu:
sudo apt install build-essential cmake libcurl4-openssl-dev
- macOS:
- Python 3 for the helper bridges (search / NVIDIA / MPC)
- At least one model backend (see Backends)
git clone https://github.com/mont127/GalliviumCloud.git
cd GalliviumCloud
make # builds ./build/ocli
make install # installs to ~/bin/ocli (+ helper scripts alongside it)Make sure ~/bin is on your PATH, then run ocli.
1. Get a model. The fastest path is the bundled default, Qwythos:
install_qwythos.sh # downloads the GGUF to ~/models and registers it with ollama(or point OCLI at any backend below).
2. Run it.
ocliYou land in the split-pane live view: type a prompt in the bar at the bottom, watch the AI answer on the left and work in the terminal on the right.
3. Pick a backend/model if needed:
/backend llama-cpp # or: ollama, openai, anthropic, nvidia, mlx
/model # arrow-select a model
4. Give it real work — it uses tools and carries the task through to a result:
scan this repo for a path-traversal bug and write a failing test that proves it
- Type in the bottom bar to chat; Enter sends.
- Ctrl-T toggles focus into the right-hand shared terminal so you can run commands yourself — the AI sees what you run, and its output can be sent along with your next message.
- Mouse wheel or PgUp/PgDn scrolls the conversation (and the terminal pane).
- The pill shows a live context-window bar (
ctx ▰▰▱▱ NN%). - Double-Ctrl-C quits.
/classicswitches to a plain scrolling view (or start withLOREA_CLASSIC=1 ocli). /terminalattaches full-screen to the shared shell (Ctrl-] to detach).
/dashboard
Opens a dark, LAN-reachable web UI (JetBrains Mono) with the same conversation plus a
live xterm.js terminal. Open the printed URL from any device on your network.
The dashboard runs arbitrary shell commands. Only use it on a trusted network.
| Backend | What it is | Setup |
|---|---|---|
llama-cpp |
Local GGUF via a llama-server |
/setup installs llama.cpp; default model = Qwythos (install_qwythos.sh) |
ollama |
Local Ollama | run ollama; install_qwythos.sh registers qwythos |
mlx |
Apple-Silicon MLX server | /setup_mlx |
openai |
OpenAI API | export OPENAI_API_KEY=... |
anthropic |
Claude API | export ANTHROPIC_API_KEY=... |
nvidia |
NVIDIA NIM (OpenAI-compatible) via a Python bridge | key in ~/.config/lorea/nvidia_api_key or NVIDIA_API_KEY |
| MPC | A remote OCLI model server you connect to | /connect <url> [--token t] |
Switch with /backend <name> and pick a model with /model. llama-cpp and ollama
default to Qwythos once the GGUF is installed.
The working context is model-aware; Qwythos gets a large window. Override it (bounded by your RAM):
LOREA_CONTEXT=128k ocli # also 256k, 1M, .../effort sets how hard the model works — and, crucially, how much it thinks before it
acts. It ranges from basic (answer directly) through elite and mythic to
beyond (deliberate to the maximum inside its reasoning before committing).
Serve any model over OCLI's MPC protocol and get a public URL, from any machine with Docker:
cd mcp-docker
docker compose up --buildOn first start it downloads/serves Qwythos and prints:
==================================================================
OCLI MPC is LIVE at: https://<name>.trycloudflare.com
Connect from any OCLI:
/connect https://<name>.trycloudflare.com --token <token>
==================================================================
Paste that /connect … line into any OCLI. To use a faster native model (your GPU)
instead of the CPU container, point the proxy at it:
MPC_UPSTREAM_URL=http://host.docker.internal:11434/v1 docker compose up --build mcpYou can also run the MPC server without Docker: mcpserver.py (needs the openai
package and an OpenAI-compatible upstream). See mcp-docker/README.md for details.
Session: /status · /usage (token + context + activity stats) · /tasks · /save [name] · /load [name] · /sessions · /exit
Workspace: /cmd <shell> (run without using a turn) · /diff [path] · /copy · /retry · /undo · /clear
View & terminal: /live · /classic · /terminal · /dashboard
Agents: /agent [n] [goal] · /agent --full [n] [goal] (agents may run commands / edit files)
Models & backends: /backend · /model · /connect [url] [--token t] · /setup (llama.cpp) · /setup_mlx · /download <repo> · /download_model [m]
Behavior: /auto (auto-execute) · /effort · /loop <goal> (autonomous; Esc stops) · /plan · /vram [--auto] · /theme [name] · /help
| Variable | Effect |
|---|---|
LOREA_CLASSIC=1 |
start in the plain scrolling view (no split-pane) |
LOREA_CONTEXT=128k |
set the working context window (128k, 256k, 1M, …) |
NVIDIA_API_KEY |
key for the nvidia backend (or use ~/.config/lorea/nvidia_api_key) |
OPENAI_API_KEY / ANTHROPIC_API_KEY |
cloud backend keys |
PYTHON |
python interpreter for the helper bridges |
BINDIR (make) |
install directory (default ~/bin) |
OCLI is a security assistant built for authorized, lawful work — your own systems, engagements you are hired for, CTFs, labs, security research, and education. Use it only where you have permission. It refuses and redirects clearly unauthorized or mass-harm requests.
See LICENSE.