Questions& answers.

An autonomous AI coding & infrastructure agent for Windows. It doesn't just advise — it runs real tools: shell / PowerShell, files, SSH, databases, the browser, image generation, vision and web search, all from one chat.

Execute shell and PowerShell commands, read and write files, SSH into servers, query MySQL / Postgres, drive the browser through MCP connectors (Chrome / Playwright / Puppeteer, GitHub / GitLab, Google Drive, Redis / Mongo), generate and recognize images, and search the web — end to end.

There are three tiers: Free ($0), Standard ($199 / month) and Pro ($499 / month) — they differ in speed, queue delay and model filtering. See the Pricing page for the full comparison.

Cursor is a cloud-connected code editor: it suggests edits inside the editor and routes your context through its cloud. NoTrack Coder is an autonomous agent, not an editor — it runs shell commands, edits files, and drives SSH, databases and the browser end to end, on your machine, with any OpenAI- or Anthropic-compatible model you connect. It runs 100% locally with no telemetry.

Yes, for a broader job. Copilot autocompletes code and chats inside your IDE using GitHub-hosted models. NoTrack Coder acts beyond the editor — it executes real tools (shell, files, SSH, databases, the browser, image generation and web search) from one chat, runs locally, and works with any model you point it at rather than a single provider.

Claude Code is a terminal agent tied to Anthropic's models. NoTrack Coder takes the same acts-not-advises approach as a Windows desktop app, but is model-agnostic: connect any OpenAI-compatible or Anthropic endpoint, or a local model via Ollama / LM Studio / llama.cpp. Chats, API keys and working folders stay 100% local, with no telemetry.

Yes. NoTrack Coder runs 100% locally on your machine — chats, working folders and API keys stay on your computer. There is no telemetry and no account is required to run it.

No. Keys, passwords and tokens that show up in tool output are redacted automatically before anything is sent to the model.

Dangerous commands are blocked outright. Risky ones ask you for confirmation first (Ask mode), and success is judged by the exit code.

You pick the approval mode per chat: Auto-run executes immediately, Ask first pauses before each state-changing tool, and Plan only plans without executing. Effort tunes how hard it reasons, and Clean mode hides the actions to show just the final answer.

The local API is guarded by a per-session token, so nothing else on your machine can drive the agent.

Any OpenAI-compatible /v1 endpoint (OpenAI, vLLM, llama.cpp, LM Studio, Ollama, OpenRouter, proxies), the Anthropic Messages API, or a local ComfyUI for images. A built-in uncensored coding model is included out of the box, and Test auto-detects each endpoint's context window.

Yes — connect any compatible model and switch as you like; the agent works with whatever you point it at.

Windows 10 / 11 (x64). The WebView2 runtime is installed automatically by the installer.

The installer is recommended: it sets up WebView2 and enables auto-updates. The portable single .exe needs no install and runs from anywhere.

Windows is available today. A Linux build is planned.

New versions are published automatically; the app shows an Update button when one is available.

Yes. With opt-in device reporting, a team sees every install in one admin panel — hardware, runtimes, installed models and online status — and can drive them from scripts via a token-authenticated Fleet API. Chat content, prompts and API keys are never sent.