AI agents. Under your command.
Contenox is built for demanding work — and the people accountable for it.
curl -fsSL https://contenox.com/install.sh | sh irm https://contenox.com/install.ps1 | iex Installs contenox. Run contenox auto to detect your hardware, download a model, and open the terminal.
No account required. Local, private, and hosted models can work side by side.
Beam reviews a local change with a model running on this machine.
The work
One standard. However you work.
At the keyboard, inside the editor, or running without you: the same work, context, and controls carry through.
The Interactive Terminal
Full-screen, native scrollback, / for commands, and @ to put a file in front of the agent. A gated call raises an approval card answered with one keystroke. Bare contenox opens it.
The Editor Integration
The same agent over stdio to Zed, JetBrains, AionUi, OpenClaw and any client speaking the Agent Client Protocol. The editor owns your workspace and permission UI.
The Automation Engine
Headless batch execution for CI/CD pipelines, cron jobs, and unattended scripts. Runs tasks with tools on that machine and exits 0 when the work lands.
Capability
The systems behind the work.
Code is only one part of real work. Contenox reaches the systems it needs without surrendering the boundaries around them.
Every call crosses one boundary and is checked against your policy envelope before it runs.
Declared per agent
Tools are passed to agents explicitly in their Markdown declarations (tools: Read, Glob, Grep, or *). Enforcing narrow toolsets prevents confusing the model with unused tool descriptions.
Added via CLI or MCP
Connect any Model Context Protocol (MCP) server or wrap internal HTTP APIs via OpenAPI specs. Declarations can also declare private, agent-scoped MCP servers.
Command
Nothing important is left implicit.
Every run is bounded by an envelope: a JSON policy naming what passes silently, what stops for a human, and what is denied outright, plus hard ceilings on tool calls and tokens. Anything no rule matches fails closed.
A run that stops for a person checkpoints where it stopped, saves the ask, and releases the process. Restart the box, close the laptop, let days pass: when the answer arrives the run resumes from that exact point, exactly once.
The Linux Landlock sandbox confines foreign agent code locally with scrubbed environments and filesystem confinement.
Run native inference through modeld, connect Ollama or vLLM, and add hosted providers with your own keys — side by side, under the same policies and execution state.
Built on
SQLite is the zero-dependency default. PostgreSQL, NATS, and Valkey form the server-backed deployment.
The Contenox Project is 100% Fair & Open Source under the Apache-2.0 License.
