Now out and live — v2.1.4
Windows PowerShell installer • zero npm dependencies

A terminal coding agent
you actually run

Omni is a from-scratch, tool-calling coding agent for the terminal. Bring your own provider — Agnes AI, NVIDIA NIM, OpenAI, OpenRouter, local llama.cpp — and it inspects, edits, tests, and ships code in your own working directory.

Get the installer View source on GitHub

No account • no telemetry • bring your own API key

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

terminal coding agent   model: nvidia/glm-5.2
type /help for commands, /exit to quit

› fix the failing test in src/parser.mjs

(o_O)? Ponderising… reading files
  ⚙ run_test npm test
  ⚙ edit_file src/parser.mjs
    (^‿^) Omi: patch applied clean.

● Fixed the off-by-one in tokenize() — tests pass.

What's actually in the box

No accounts, no cloud sync, no hidden services. Everything below runs on your machine.

Bring your own provider

NVIDIA NIM (free default) or Agnes AI (also free), plus OpenAI, OpenRouter, Groq, DeepSeek, Google, xAI, Mistral, or a bundled local llama.cpp server. Dual-account failover switches keys automatically on a rate limit.

Real tool-calling loop

Reads and patches files, runs shell commands and tests, inspects git, manages long-running dev processes — all gated by a per-tool permission system you control.

Workspace sandboxing

File tools are confined to a trusted folder by default, with symlink-aware containment. Widen it deliberately with /workspace scope system when you mean it.

Semantic code intelligence

LSP-backed definition/reference/hover lookups, symbol search, dependency auditing, test-coverage detection, and a built-in security scanner.

Self-extending

create_tool lets the agent write and hot-load a new tool into the running session — no restart. Skills, MCP servers, and installable packages extend it further.

Zero npm dependencies

Pure Node ≥ 20 and the built-in fetch. Nothing to audit in node_modules because there isn't one.

/neuralview — live, in your browser

Watch its memory think

A rotatable 3D map of how Omni actually remembers things — rendered from a plain node:http server and hand-rolled canvas 3D, no WebGL, no Three.js, nothing external. This is the real graph, not a mockup.

  • Memory Coordinator — the center. Every retrieval and every captured event routes through here.
  • Chat Memory — preferences, facts, and decisions are weighed and placed automatically. No manual add, no approval step; a heavier, more-recent fact simply supersedes an older one.
  • Skills & Wiki — your OKF knowledge base, split by card type: reusable howto/pattern procedures vs reference and decision cards.
  • L0 & Feedback — the immutable session log, and the loop that resolves conflicting memories by weight.
inside Omni
/neuralview
Omni Neural Network — a live 3D graph of the Memory Coordinator, L0 conversation log, Chat Memory atoms, OKF Skills and Wiki, and the Feedback loop, updating in real time as the agent works

Install in one line

Checks the latest release, downloads it, and wires up the omni command globally.

PowerShell
irm https://omni.globalwarningnetworks.com/install.ps1 | iex

Requirements

  • • Windows 10/11 with PowerShell
  • • Node.js 20 or newer
  • • A free NVIDIA NIM or Agnes AI key, or any OpenAI-compatible provider

Agnes AI's free Starter tier: 1,500 agnes-2.5-flash requests per 5 hours.

Re-run the same command anytime — it skips straight to "already up to date" when there's nothing new.

Then cd into a project, run omni, and type /neuralview to watch its memory live — see it above.

Your keys never leave your machine

The installer copies .env.example to .env and leaves it empty — no keys are bundled, generated, or fetched for you. Set your own with /apikey <provider> <key> from the REPL, or edit .env directly.