Codex
Architecture, implementation, testing, review and Git workflows.
NEM0W.
A local-first Windows desktop system for turning a software objective into coordinated, reviewable work across AI coding agents.
02 / PURPOSE
Why it exists.
nem0w is a personal development environment for Windows. It combines a private ChatGPT conversation space with a project room where several authenticated coding assistants can work on the same objective without competing for the same files.
The application plans responsibilities, gives each agent an isolated Git branch and worktree, passes real output and code between stages, and keeps the final merge away from the main branch until it has been reviewed. The goal is not to hide the work behind a single prompt; it is to make the work visible, interruptible and accountable.
It is still in active development. The current code already covers chat, project orchestration, staged squads, local memory, encrypted persistence, recovery questions, review and controlled delivery.
03 / CONNECTIONS
AI clients it can coordinate.
nem0w connects to official command-line clients already authenticated on the computer. It does not ask for or store their API keys.
Claude Code
Large-context work, architecture, documentation and code review.
Antigravity
Frontend, multimodal, research and testing tasks.
Copilot CLI
Implementation, GitHub workflows, tests and review.
Cursor Agent
Implementation, refactoring, debugging and frontend work.
Grok Build
Research, debugging and implementation tasks.
Vibe
Terminal work, implementation and code review.
Auggie
Large-codebase context, implementation and review.
Sourcegraph Amp
Implementation, debugging and review from the terminal.
DRAG TO SPIN · OR USE ← →
04 / WORKFLOW
From objective to reviewed code.
- 01PLAN
The objective becomes a technical blueprint with roles, stack decisions, file ownership and contracts.
- 02ISOLATE
Each agent receives its own Git branch and worktree, protecting the main checkout while work is in progress.
- 03HAND OFF
Sequential stages receive both the previous agent's explanation and the files that explanation refers to.
- 04REVIEW
Changes are collected in a separate integration branch where diffs and conflicts can be inspected.
- 05DELIVER
The main branch changes only after an explicit review and delivery action.
05 / CAPABILITIES
What it does.
- 01Personal ChatGPT conversations with selectable account models, searchable local history and optional voice dictation through Windows
- 02Project planning with technical blueprints, assigned responsibilities, file ownership and execution stages
- 03Ready-made squads for a sequential four-agent pipeline, website work and full-stack applications, plus custom teams of up to six agents
- 04Parallel or staged execution through official coding CLIs, with live logs, usage data, pause, resume, retry, cancellation and agent replacement
- 05Persistent handoffs and project memory so later agents receive the decisions and implementation produced earlier
- 06Human-in-the-loop questions: blocked agents can ask for a decision and continue after the answer instead of losing their work
- 07Review panels for changed files and diffs, followed by isolated integration and an explicit final delivery to the main branch
- 08Local encrypted storage for conversations, memories and project state, with secret filtering and password-protected export and restore
06 / TECHNOLOGY
Under the hood.
- 01C# and .NET 8 — the application logic, process coordination, persistence and Windows integration
- 02WPF and XAML — a native Windows interface with reusable resources, custom controls and per-monitor DPI support
- 03Git worktrees and branches — physical isolation for every agent, staged integration, diff review and controlled delivery
- 04Official AI coding CLIs — adapters for Codex, Claude Code, Google Antigravity, GitHub Copilot CLI, Cursor Agent, Grok Build, Mistral Vibe, Augment Auggie and Sourcegraph Amp
- 05Codex App Server — account login, model discovery, usage limits and persistent ChatGPT conversation threads without storing an API key
- 06Model Context Protocol — an embedded MCP server built with ASP.NET Core, Kestrel and ModelContextProtocol.AspNetCore so an agent can ask the user a question during execution
- 07Windows DPAPI and JSON — user-bound encryption for local conversations, memories and project records, written atomically to disk
- 08AES-GCM and PBKDF2-SHA256 — authenticated, password-derived encryption for portable backup files
- 09Windows native APIs — process management and SendInput integration for the system voice-typing panel
- 10Custom verification project — checks for encryption, Git isolation, handoffs, staged execution, backups, encoding and failure recovery
07 / ENGINEERING
Safety by architecture.
nem0w removes known API-key and alternate-endpoint variables before launching supported agents. The intended connection is through each official CLI and its own account login, keeping authentication outside the application's data store.
Agent output is treated as untrusted data when it enters shared memory. Potential secrets are rejected, context blocks are delimited, events are capped, and project state is preserved when work is paused or an agent fails.
Destructive Git actions are deliberately separated from ordinary execution. Agents work outside the main checkout; integration happens on another branch; delivery to main is a distinct user-approved step.
nem0w is my attempt to turn AI-assisted development into a process I can inspect and control: clear responsibilities, isolated code, durable context and a final decision that still belongs to the developer.