Open source · AGPL-3.0 · runs in your terminal

███╗   ██╗██╗   ██╗██╗  ██╗████████╗██╗  ██╗ ██████╗ ███████╗
████╗  ██║╚██╗ ██╔╝╚██╗██╔╝╚══██╔══╝██║  ██║██╔═══██╗██╔════╝
██╔██╗ ██║ ╚████╔╝  ╚███╔╝    ██║   ███████║██║   ██║███████╗
██║╚██╗██║  ╚██╔╝   ██╔██╗    ██║   ██╔══██║██║   ██║╚════██║
██║ ╚████║   ██║   ██╔╝ ██╗   ██║   ██║  ██║╚██████╔╝███████║
╚═╝  ╚═══╝   ╚═╝   ╚═╝  ╚═╝   ╚═╝   ╚═╝  ╚═╝ ╚═════╝ ╚══════╝

The most customizable open-source agentic coding harness.

Bring your own model — any open model, any host. Plug in an API key and go. Five safety-first systems keep long agent runs on the rails, so you can hand off real work without holding your breath.

View source ↗
$ npm install -g nyxthos click to copy

Five systems, one harness

Built for long runs that don't fall apart.

Most coding agents are one prompt away from wandering off, looping, or shredding a file. nyxthos wraps the model in five independent systems that hold the line.

codegraph

Reads code as a graph

Real ASTs — classes, methods, dependency edges — indexed so the agent asks for the function it needs instead of grepping blind through text.

sentinel

Kills doom-loops

Three identical actions and it hard-freezes, wipes the loop from memory, and forces a brand-new strategy — no more watching an agent bang the same wall.

shadow git

Nothing is unrecoverable

Every write is a micro-commit on a hidden branch. Visual diffs, one-command rollback. The agent can't destroy what a keystroke can restore.

context engine

Spends tokens like they cost money

Sends only what matters — changed files, error traces, the task log — and prunes the rest before it ever reaches your budget.

sandbox

Zero-config isolation

Tool commands run confined to the workspace. API keys and secrets are scrubbed from the environment and brokered at spawn — never in a log, never in a transcript.

No vendor lock-in

Bring your own model.

nyxthos never drives one company's model for you. Point it at a local runtime or a hosted endpoint, drop in an API key, and it speaks the protocol — native function-calling when the model supports it, a tagged-text fallback when it doesn't.

Ollama vLLM LM Studio llama.cpp OpenRouter Groq Together Fireworks DeepInfra Hugging Face Replicate RunPod Anyscale Lambda any OpenAI-compatible

Get started

One command, then nyxthos.

Install the CLI, run it in any project directory, and it opens straight into a fullscreen terminal app. On first run it walks you through choosing a model provider.

bash — install
$ curl -fsSL https://nyxthos.com/install.sh | sh
# checks for Node 20+, installs the CLI — that's it
# — or with your package manager —
$ npm install -g nyxthos
$ nyxthos # opens the TUI in the current directory
# — or build from source —
$ git clone GITHUB_URL && cd nyxthos && pnpm install && pnpm build