orchestrator · not an agent

AI Coding Agent
Orchestrator

— loops until done.

Give it a spec. It writes tests, implements code, verifies quality, and retries until everything passes.

GitHub
plan
acceptance
route
execute
verify
review
escalate
loop
regression

Why nax

Built around discipline,
not convenience.

AGENT-AGNOSTIC

Works with any agent

Claude Code, Codex, Gemini CLI, or any ACP-compatible agent. nax drives them all through the same disciplined loop.

TDD-ENFORCED

Tests first, always

Acceptance tests must be RED before implementation starts. No shortcuts, no skipped steps — the loop enforces it.

LOOP UNTIL DONE

Verify, retry, escalate

Failed verify? Retry. Still failing? Escalate to a more powerful model tier. Loop until every test passes.

MONOREPO-READY

Per-package everything

Per-package config, per-package test commands, per-story working directories. Full monorepo isolation out of the box.

SEMANTIC REVIEW

LLM-powered quality gates

Behavioral review against story acceptance criteria. Catches stubs, placeholders, and out-of-scope changes before they merge.

ADVERSARIAL REVIEW

Probes edge cases

A second LLM pass that attacks the implementation — probing input handling, error paths, and abandoned implementations.

LANGUAGE-AWARE

Auto-detects your stack

Detects Go, Rust, Python, TypeScript from manifest files. Adapts test commands, mocking patterns, and context per language.

EXTENSIBLE

Plugin system

Hook into routing, review, reporting, and post-run actions. Drop a plugin into .nax/plugins/ and it is picked up automatically.

How it works

A loop with teeth.

01

Plan

optional

Generate prd.json from a spec file using an LLM. Stories are decomposed, sized, and sequenced automatically.

02

Acceptance setup

Generate acceptance tests and assert RED before implementation begins. No green tests allowed at this stage.

03

Route

Classify story complexity and select a model tier — fast, balanced, or powerful. Security-critical stories always escalate.

04

Execute

Run the agent session. Claude Code, Codex, Gemini CLI, or any ACP-compatible agent. Context and constitution are injected per story.

05

Verify

Run scoped tests. On failure, the rectification loop fires first — fixing before escalating to the next model tier.

06

Review

Lint, typecheck, semantic review against ACs, and adversarial review for edge cases. Autofix runs before escalation.

07

Escalate & loop

On repeated failure, retry with a higher model tier. Repeat until all stories pass or cost / iteration limits are hit.

08

Regression gate

Full test suite runs once after all stories complete. nax retries failed suites; a gate failure does not block the feature.

09

Acceptance

Run the original acceptance tests against the finished feature. GREEN here means done.

Agents

Bring your own agent.

nax doesn't care which model you use. Swap agents without touching your workflow.

recommended

Claude Code

ACP · CLI

Full ACP support with multi-turn continuity, token-cost tracking, and structured JSON-RPC. Recommended.

Codex / OpenCode

CLI

OpenAI Codex CLI adapter. Pass --agent opencode to activate.

Gemini CLI

CLI

Google's Gemini CLI adapter. Pass --agent gemini to activate.

Any ACP agent

ACP

Any agent implementing the Agent Client Protocol connects via acpx. Structured sessions, cost alignment, multi-session continuity.

Install

Requires Bun 1.3.7+ or Node 22+.
Git must be initialized in your project.

npm

$ npm install -g @nathapp/nax

bun

$ bun install -g @nathapp/nax

Quick start

cd your-project
nax init                           # Create .nax/ structure
nax features create my-feature     # Scaffold a feature

# Write your spec, then run
nax run -f my-feature --plan --from spec.md

Or skip interactive planning with --plan --from spec.md for a single-command run.