Local AI agent CLI

Describe a goal.
Get it done.

AI workflows at your fingertips.

Contenox turns natural language into persistent, step-by-step plans and executes them on your machine using real shell and filesystem tools.
No cloud required. No daemon. Just a single binary.

$ contenox plan new "install a git pre-commit hook that blocks commits when go build fails"
Creating plan "install-a-git-pre-commit-a3f9e12b" — 5 steps

$ contenox plan next --auto
Step 1: Install necessary tools... ✓
Step 2: Create .git/hooks/pre-commit... ✓
Step 3: Write the build-check script... ✓
Step 4: Write bash content to hook file... ✓
Step 5: chmod +x .git/hooks/pre-commit... ✓

# The model wrote that hook. On your machine.
Get Contenox GitHub
  • 🏠 Fully local with Ollama
  • 💾 Plans persist across reboots
  • 🔍 Review before every step
  • 🔌 OpenAI, Gemini, vLLM too

Not another chatbot wrapper

Most AI CLI tools are one-shot: ask a question, get a reply, done. Contenox is different because it's built on a typed execution engine — not a prompt loop.

Other tools

One-shot, stateless, forgotten

Copilot CLI suggests a command. Cursor autocompletes code. Mistral CLI hits the cloud API. None of them remember what happened, can retry a failed step, or execute a 5-part plan autonomously while you grab a coffee.

Contenox

Persistent plans, real execution

Plans survive reboots. Each step result is saved to SQLite. You can pause mid-plan, inspect what happened, retry a specific step, or let the model replan from the current state. It's a workflow engine you run from a terminal.

Real execution engine under the hood

Contenox runs a typed state machine: branching logic, retry policies, multi-model routing, tool call dispatch, token budget management. The CLI surface is simple. What's underneath isn't.

Human-in-the-loop by default

contenox plan next runs one step and pauses. You review. You decide. --auto unlocks full autonomy — only when you say so. This isn't a safety feature bolted on. It's the design.

Your chains, your models

Workflows are JSON files you own. Swap models per task. Add external hooks for anything that has an HTTP API. Run fully offline with Ollama. Or use OpenAI, Gemini, or any OpenAI-compatible endpoint.

Zero infrastructure

Single binary. SQLite. No daemon, no Docker, no NATS, no Postgres. Drop it on any machine and it works. The Runtime API server is available for team and production use — but the CLI needs nothing.

Pure UNIX.

Contenox is an autonomous engine, not a chatbot wrapper. There is no TUI or GUI to trap you. It uses standard streams, exits when done, and is designed to run headless in CI/CD pipelines or be composed with other UNIX tools.

Install

One command. Move the binary. Done.

# Download latest release
TAG=$(curl -sL https://api.github.com/repos/contenox/contenox/releases/latest | grep '"tag_name"' | cut -d'"' -f4)
ARCH=$(uname -m | sed 's/x86_64/amd64/;s/aarch64/arm64/')
curl -sL "https://github.com/contenox/contenox/releases/download/${TAG}/contenox-${TAG}-linux-${ARCH}" -o contenox
chmod +x contenox && sudo mv contenox /usr/local/bin/contenox
$ contenox --version
# Download latest release
TAG=$(curl -sL https://api.github.com/repos/contenox/contenox/releases/latest | grep '"tag_name"' | cut -d'"' -f4)
ARCH=$(uname -m | sed 's/x86_64/amd64/;s/arm64/arm64/')
curl -sL "https://github.com/contenox/contenox/releases/download/${TAG}/contenox-${TAG}-darwin-${ARCH}" -o contenox
chmod +x contenox && sudo mv contenox /usr/local/bin/contenox
$ contenox --version
git clone https://github.com/contenox/contenox
cd contenox
go build -o contenox ./cmd/contenox
sudo mv contenox /usr/local/bin/contenox

Then scaffold your config and run your first task:

# Local model
ollama serve && ollama pull qwen2.5:7b

$ contenox init
Created .contenox/ with default chains...

# Chat with your model
$ contenox "hey, what can you do?"

# Plan and execute a multi-step task
$ contenox plan new "create a TODOS.md from all TODO comments in the codebase"
$ contenox plan next --auto

Cloud providers (OpenAI, Gemini, vLLM) also work — contenox backend add openai --type openai --api-key-env OPENAI_API_KEY

How it works

Three modes. One binary.

contenox plan — Autonomous task execution

Describe a goal. The LLM generates a plan. Steps are saved to SQLite.
Execute one step at a time with plan next, or go full-auto with plan next --auto. Retry, skip, or replan any step at any time.

contenox chat — Interactive chat

Natural language → shell commands → response. Chat history persists across sessions. Use your own chain with --chain, or use the default.

contenox run — Scriptable chains

Run any chain with any input type, stateless. Pipe-friendly. Perfect for CI scripting, batch jobs, or connecting chains together.

Bring your own workflow

Chains are JSON. Define tasks, handlers, branching conditions, tool hooks, and model configs. Swap models per step. Add any HTTP endpoint as a tool. The engine handles the rest.

Native MCP client — AI with real memory

Contenox speaks the Model Context Protocol — the open standard donated by Anthropic to the Linux Foundation that lets agents form lasting connections to real tools.

Most agents forget the moment a tool call ends. Contenox maintains a persistent, session-scoped connection to every registered MCP server. Your agent picks up exactly where it left off — across tool calls, plan steps, even restarts.

# Register a local filesystem MCP server
$ contenox mcp add myfiles --transport stdio \
  --command npx --args "-y,@modelcontextprotocol/server-filesystem,/home/user/projects"

# Or a remote HTTP server
$ contenox mcp add memory --transport http --url http://mcp-host:8090

# Now run a plan — the agent carries context across every step
$ contenox plan new "explore codebase, identify top 3 bugs, write a report"
$ contenox plan next --auto

True session memory

Each plan or chat gets its own persistent connection. State survives tool calls, thinking steps, and full plan runs.

Any transport

stdio for local processes, sse for event-stream servers, http for production deployments. All feel native.

Any MCP server

Filesystem, databases, GitHub, internal company tools — if it speaks MCP, Contenox connects to it instantly.

Same CLI, same server

Everything you build locally works identically in the production Runtime API. Zero config changes needed.

Read the MCP integration docs →

Ready to try it?

One binary. Works on your laptop, your server, or air-gapped.

Get Contenox Star on GitHub ⭐

Questions: hello@contenox.com