Noeza — Think. Build. Secure.

The Noeza AI coding agent for VS Code, by NT-R LTD. Free.

Noeza brings an autonomous coding agent into a VS Code side panel — it reads your project, plans work, proposes reviewed diffs, and runs commands in your own visible terminal, only with your approval. You bring your own model; Noeza brings the agent, the safety, and the transparency.

Two things make Noeza different:

• Every command runs in your terminal, in the open. When the agent builds, tests, or runs a tool, it does so in your real, visible terminal — the same one you use — so you see the exact command and its output as they happen, and can interrupt it. Not a replay of a hidden process: the terminal doing the work.
• You bring your own model. Noeza runs none of its own, and NT-R runs no model in the path. Connect an external agent (Claude Code), a frontier provider under your own key, or any OpenAI-compatible endpoint you host — with one /connect command. Your key and your code go to the model you chose; they never pass through NT-R.

Install

From your terminal:

curl -fsSL https://noeza.ntr-kenya.org/install.sh | sh

It downloads one signed .vsix, verifies its checksum, and installs it with your own code CLI — nothing hidden. Prefer to do it by hand:

curl -fsSL https://noeza.ntr-kenya.org/noeza.vsix -o noeza.vsix
code --install-extension noeza.vsix

Requires a VS Code-family editor (VS Code 1.93+, VSCodium, Cursor) with the code command on your PATH.

First run

1. Press Ctrl/Cmd + Shift + A to open the Noeza panel (or click the Noeza icon in the Activity Bar).

2. Connect a model you already trust:

   /connect <base-url> <model> [api-key]
   

   any OpenAI-compatible endpoint — a frontier provider under your own key, or a self-hosted one (vLLM, llama.cpp, LM Studio) — or run Claude Code inside the editor. Your API key is stored in the editor's secret storage, never in a settings file, and is only ever sent to the endpoint you set. /connect off disconnects and clears it.

   Keys are stored one per provider in the OS keychain, so connecting to a second provider never overwrites the first — switch back and forth without re-pasting. /keys lists what is held and where each key came from.

   To keep a key across reinstalls, set it in your shell profile. The environment lives outside Noeza's install directory and profile, so an upgrade cannot lose it:

   export NOEZA_QWEN_API_KEY="sk-ws-..."
   export NOEZA_OPENAI_API_KEY="sk-..."
   export NOEZA_ANTHROPIC_API_KEY="sk-ant-..."
   

   Noeza reads its own NOEZA_* variables rather than a provider's conventional name (OPENAI_API_KEY): adopting a variable you set for something else would send that credential to whatever endpoint Noeza happened to be pointed at. A key in the keychain wins over one in the environment.

   A key is only ever sent to the provider it was stored for. Two consequences worth knowing:

   • A local runtime never receives a key. Ollama and similar ignore the value, so there is nothing to gain and a credential to lose.
   • A self-hosted or unlisted endpoint keeps its key per host, set with /connect <url> <model> <key>, and is never filled in from the environment — nothing about such a URL can be verified, so the key has to be one you typed for that exact endpoint.

3. Ask for something: /plan add rate limiting to the login route.

Two Noeza products, one free plan

They are independent — neither needs the other, both free:

See https://noeza.ntr-kenya.org/docs.

What that means for your code

The editor runs on your machine: editing, indexing, permissions, and the audit log all stay on-device. NT-R runs no model, so an agent request's code and context go only to the model you connect, under that destination's policy — never to NT-R. Point Noeza at a self-hosted or local endpoint and an agent request stays on your own machine too. NT-R never receives, stores, or proxies the API key you configure.

Commands

Type / in the panel for autocomplete. Highlights:

You can also right-click a selection → Noeza: Explain / Edit Selection, or press Ctrl/Cmd+Shift+A to focus the panel.

Permission modes (security-first)

Risky commands (rm -rf, git push, curl | sh, sudo, …) are flagged High and always require approval, even in auto-all. Every command runs in your visible terminal regardless of mode — you always see what ran.

How execution works

When the agent runs a command, Noeza runs it in the editor's real integrated terminal and reads its output and exit code from there — the same terminal you type into. There is no hidden subprocess: if shell integration is available it captures output directly; if not, the command still runs visibly (piped through tee) so you watch it live. A timeout or your stop sends the same Ctrl+C you would.

Architecture

src/
  extension.ts          Activation, commands, panel registration
  panel/provider.ts     WebviewViewProvider — bridges UI ↔ agent, routes slash commands, /connect
  agent/
    loop.ts             Agent orchestration loop (model turns → tools → approvals)
    tools.ts            read/list/search/write/edit/run/git tools
    terminal.ts         Visible integrated-terminal runner (no hidden subprocess)
    permissions.ts      Permission engine + command risk assessment
    prompts.ts          System prompt + per-command prompts
    audit.ts            Append-only audit log (.noeza/audit/, secrets redacted)
  model/
    client.ts           OpenAI-compatible streaming client (+ tool calls)
    endpoint.ts         Endpoint resolution — never invents a default (unconfigured → /connect)
  context/
    workspace.ts        Safe paths, ignore rules, secret redaction
    collector.ts        Live editor + repo context
  commands/registry.ts  Slash-command registry + parser
media/                  Webview UI (main.js, main.css, icons)

The agent loop mirrors the Claude Code / Codex pattern: the model streams text and tool calls; reads run automatically; writes and commands pass through the permission engine (auto-allow / ask / deny) before executing; results feed back until the model finishes or hits the step limit.

Security & robustness built in:

• Prompt-injection defense — all repo-derived content (file reads, search results, git diffs, selections) is wrapped as untrusted data so a hostile repo can't hijack the agent.
• Undoable edits — writes apply through VS Code WorkspaceEdit, so every change is visible in the editor and reversible with Ctrl/Cmd+Z.
• Visible execution — commands run in your real terminal; nothing runs where you can't see it.
• Context budgeting — conversation is trimmed to a token budget; /compact summarizes and resets. Auto-compaction keeps long sessions working.
• Audit + redaction — actions logged to .noeza/audit/; secrets redacted from context and logs; .env/key files refused.

Build from source

cd extension
npm install
npm run compile     # dev build
npm run typecheck   # tsc --noEmit
npm test            # unit tests
npm run package     # vsce package → .vsix (then scan-package verifies the artifact)

A production build (npm run build) ships with no default endpoint — the user connects one with /connect. npm run package scans the built .vsix and refuses to produce one that carries a localhost default or a leaked secret.

License & attribution

© 2026 NT-R LTD. Noeza is proprietary, licensed under the Noeza Commercial Licence (see LICENSE). The bundled media/icon.png is a generated placeholder — replace with brand art before a marketplace listing.