AI coding tools
don't think in projects.

Most AI development tools operate at the file or prompt level. They generate code but don't track what was built, why it was built, or what comes next. The result: fast output, slow progress, and codebases that accumulate debt faster than teams can address it.



n-dx operates at the project level. A living PRD drives every cycle, so work is structured, traceable, and repeatable — regardless of which model or provider is doing the execution.

Vibe-coded apps with no structure Fast iteration without a spec creates codebases that are difficult to reason about, maintain, or hand off. Pattern and anti-pattern detection surfaces the full picture.
🗺
Onboarding to an unfamiliar codebase Understanding file inventory, import graphs, and architectural zones takes days without tooling. sourcevision produces a machine-readable map in minutes.
🔁
No memory between AI work sessions Without a shared source of truth, every AI session starts from scratch. rex maintains a structured PRD across every cycle, keeping state between runs.

A self-cleaning loop from
analysis to execution.

Run the loop once to get started. Run it again to stay current. The PRD reflects the codebase at every step — no manual tracking required.

1
Analyze

sourcevision scans your codebase — files, imports, zones, components — and produces a machine-readable context that all downstream tools consume. Run any time the codebase changes.

2
Plan

rex reads the analysis and generates a structured PRD. Review proposals, accept them, and maintain a live task tree with epics, features, dependencies, and acceptance criteria — updated each cycle.

3
Execute

hench picks the next task, builds a brief with full codebase context, and ships it — writing code, running tests, committing results. Mark complete. Pick the next. Repeat.

terminal n-dx loop
$ npx n-dx init .
sourcevision initialized
rex initialized
hench initialized
 
$ ndx analyze .
Analyzing codebase...
142 files · 12 zones · 38 components
 
$ ndx recommend .
Generated 6 epics, 18 tasks
PRD saved to .rex/prd.json
 
$ ndx work .
Picking next task...
► "Add user authentication"
Writing code · Running tests · Committing
✓ Task complete — 3 files changed, all tests pass
 
$ ndx status .
Epic: Auth System [▓▓▓▓░░░░░░] 40%
Epic: Dashboard [▓▓░░░░░░░░] 20%
Next: "OAuth provider integration"

Where n-dx fits in
the development workflow.

Built for teams working on real codebases — not greenfield prototypes. Each scenario below represents a common inflection point where structured tooling returns measurable value.

onboarding
Codebase onboarding
Understand an unfamiliar repo in minutes, not days.

sourcevision maps file inventory, import graphs, and architectural zones automatically. New contributors get a machine-readable picture of the full codebase on first run — without reading every file.

cleanup
Vibe-coded app cleanup
Bring structure to a codebase built for speed, not scale.

When a prototype becomes a product, code sprawl follows. n-dx detects patterns and anti-patterns across zones, generates a prioritized cleanup backlog in the PRD, and executes fixes autonomously with guardrails.

spec-driven
Spec-driven development
Let the PRD drive execution, not just document it.

Describe what needs to be built in plain language. rex structures it into epics, features, acceptance criteria, and task dependencies. hench executes from that structure — every run traceable back to the spec.

change management
Ongoing change management
Keep the plan in sync with a codebase that never stops changing.

Re-run ndx recommend any time the codebase shifts. The PRD self-updates: completed items are marked, new work surfaces, dependencies are re-validated. The loop stays current without manual tracking.

For the practitioner

Run it. Ship it.
Let the spec keep score.

n-dx is a CLI toolkit that fits into existing workflows. No new IDE, no GUI required. Point it at a directory, run two commands, and the loop starts. The PRD lives next to the code in version control.

  • ndx analyze . && ndx recommend . && ndx work . — the full workflow in three commands
  • Works with Claude API, Claude CLI, and Codex — switch providers without changing workflow
  • Every run is logged with full transcripts, token usage, and file diffs
  • PRD state in .rex/prd.json — version controlled, diff-able, readable
  • Natural language task input: ndx recommend --add "user auth with OAuth"
For the team lead

Structured output.
Traceable progress.

n-dx gives engineering teams a shared source of truth that doesn't require a separate planning tool. The PRD is generated from the codebase and stays in sync — automatically. Progress is visible without status meetings.

  • Epics, features, and tasks generated from real codebase structure — not top-down invention
  • Notion sync keeps the PRD available to non-technical stakeholders without manual updates
  • CI integration: ndx ci runs analysis and validates PRD health on every push
  • Open source, zero per-seat cost — bring your own LLM API key or use the CLI
  • Token budgets and guardrails configurable per project — cost stays predictable

The only tool that owns
the full loop.

Most AI development tools address one layer of the problem. n-dx connects all three — and keeps them in sync.

Tool n-dx Claude Code Aider OpenHands Goose MetaGPT 8090 Devin
Codebase analysis
Structured PRD / task tree
Autonomous execution loop
Vendor-neutral (multi-LLM)
Persistent state across sessions
Open source
CLI-first / local

● Full  ◐ Partial  ○ No  · Based on public documentation as of Q1 2026.

One command to initialize.
Two to run the loop.

Requires Node.js 18+ and pnpm 10.x. Works with Claude API, Claude CLI, or Codex — configure your provider on first run.

npx n-dx init .

Questions? Let's talk.

Whether you're evaluating n-dx for your team or want to discuss a specific use case — we'd love to hear from you.

Get in Touch