GitHub - navox-labs/agents: AI agents for Claude Code — 15 specialists that run your full software development sprint cycle. Strategy, architecture, code review, QA, security audit, and shipping. Open-source Claude Code plugin.

15 AI specialists. One sprint cycle. Zero dependencies. Think. Plan. Build. Review. Test. Ship. Reflect.

One plugin. 15 agents. Full sprint cycle. Install in 30 seconds — no dependencies, no platform, no login.

/plugin marketplace add https://github.com/navox-labs/agents

A Claude Code plugin and a Python CLI for autonomous orchestration. Install the plugin for interactive sprints. Use the SDK to run your team while you sleep.

Why I built this

When I heard AI will replace engineers, I wanted to find out how. Not in theory — in practice. How does a non-human replace humans in an engineering capacity?

I'm Nahrin, founder and builder at Navox Labs. Earlier in my life I worked in building architecture — reviewing blueprints, designing structures, making sure everything held together before a single brick was laid. I'm the only architect I know who transitioned to full-time software engineering without a formal CS education. I told myself: once an architect, always an architect. I understood tech through a builder's lens — scalability, optimization, reliable structure.

I've been building physical products for 16 years and tech products for 7. Twenty-three years of building, and I never thought AI could replace my ability to make decisions. It can't. But it can execute them.

Navox Agents is how I ship.

As a solo founder, I needed a reliable team that requires minimum management. Not a chatbot I have to babysit. Not a prompt I cross my fingers on. A team of specialists with defined roles, handoff contracts, quality gates, and the discipline to self-evaluate before passing work downstream.

I believe a single builder with the right tooling can move faster than a traditional team. This is my answer.

Fully open source. A Claude Code plugin for interactive sprints and a Python CLI for autonomous orchestration. I combined these two because I want this to reach as many people as possible. These agents are built with best practices from Anthropic themselves. They self-evaluate for high reliability. They're orchestrated and work exactly like a team of specialists.

This is an open-source software factory. I use it every day. I'm sharing it because these tools should be available to everyone.

Fork it. Improve it. Make it yours.

And if you want to hate on free open-source software — you're welcome to, but I'd rather you just try it first.

Who is this for

Founders and CEOs — especially technical ones who still want to ship. Run a full sprint while you focus on strategy, fundraising, or sleep.

First-time Claude Code users — structured roles instead of a blank prompt. Instead of figuring out what to ask Claude, just tell the team what to build.

Tech leads and staff engineers — orchestrated specialist agents that follow your architecture decisions, not fight them. Sprint chains enforce the same discipline you'd expect from a senior team.

Solo builders — one person, 15 specialists. The math works.

What happens when you run it

$ /agency-run FULL Build a cookie clicker game with crab theme

→ Strategist: "Single HTML file is smart — no build step, instant deploy.
   But cookie clickers need prestige mechanics to retain players. Adding to spec."
→ Spec Writer: 7-section spec with 12 acceptance criteria
→ Architect: Component tree, state machine, CSS custom properties for theming
→ Gate 1: You review. You approve.
→ UX + Security: Running in parallel...
→ Full Stack: 1,330 lines written. All acceptance criteria met.
→ Checkpoint: App running at localhost:3000 — LGTM / FEEDBACK / STOP?
→ Reviewer: 7 specialists. 2 issues found. Auto-fixed.
→ QA: 6 bugs caught. 6 bugs fixed. 14 tests passing.
→ Security: No vulnerabilities. Auth N/A. Cleared.
→ Shipper: v1.0.0 tagged. PR #1 created.
→ Retro: 4 learnings captured to project memory.

Total time: 7 minutes. Human edits: 0.

See it work

nom.sh — one prompt, 7 minutes:

A crab cookie clicker. 1,330 lines. 6 bugs caught by QA. Zero human edits. See the code

PipeWar — built, debugged, and deployed by agents:

A Factorio-inspired tower defense game. Built from scratch, 8 production bugs diagnosed and fixed, 65 tests passing. All by the agent team. See the code

Install as a Claude Code plugin

If you hit an SSH error, run this first (one time):

git config --global url."https://github.com/".insteadOf "git@github.com:"

Then install:

/plugin marketplace add https://github.com/navox-labs/agents
/plugin install navox-agents
/reload-plugins

Note: Plugin commands are namespaced. Use /navox-agents:agency-run and /navox-agents:hire-team instead of /agency-run and /hire-team. If you installed via the manual method below, no namespace is needed.

If Navox Agents saves you time, give it a star. It helps other builders find it.

Alternative: manual install (for customization)

git clone https://github.com/navox-labs/agents.git
cd agents
bash scripts/setup.sh

Options:

bash scripts/setup.sh --global             # Install to home directory (all projects)
bash scripts/setup.sh --agents strategist,reviewer  # Install specific agents only
bash scripts/setup.sh --list               # See all available agents

Or copy manually:

mkdir -p ~/.claude/agents ~/.claude/commands
cp -r .claude/agents/* ~/.claude/agents/
cp -r .claude/commands/* ~/.claude/commands/
cp ETHOS.md ~/.claude/ETHOS.md

The sprint

Three modes. Pick one based on what you need.

Full sprint — idea to shipped PR with retrospective

/agency-run FULL Build a {browser-based} {Cookie Clicker game}
with {Atari pixel art} vibes where {crabs eat cookies}.
No authentication. No backend. Single HTML file.

Strategist challenges your assumptions. Spec Writer turns it into a buildable spec. Architect designs the system. UX maps every screen. Security audits the design. Full Stack builds it. Local Review shows it to you. Reviewer runs a 7-specialist army over the code. QA finds every edge case. Security audits the code. Shipper creates the PR. Retro captures what the team learned.

Quick sprint — skip strategy, get to code faster

/agency-run QUICK Add a {dark mode toggle} to the settings page

Hotfix — bug to fix to ship

/agency-run HOTFIX Users get 403 errors after login on mobile Safari

Power tools — use agents directly

Replace the {variables} with your own idea. Plugin users: prefix with navox-agents: (e.g. /navox-agents:strategist DIAGNOSE)

How it works — FULL Sprint

flowchart TD
    CLI["`**$ /agency-run FULL** your task here`"]

    CLI -->|your task| STRAT

    STRAT["Strategist · Opus 4.6\nDIAGNOSE\nforcing questions · risk assessment · scope"]

    STRAT -->|strategic brief| SPEC

    SPEC["Spec Writer · Sonnet 4.6\nWRITE\n7-section spec · acceptance criteria · edge cases"]

    SPEC -->|buildable spec| ARCH

    ARCH["Architect · Opus 4.6\nDESIGN\nstack · auth · API contracts · security · testing"]

    ARCH -->|system design| G1

    G1{{"Gate 1\nyou review + approve"}}

    G1 -->|UX brief| UX
    G1 -->|auth model| SEC1

    subgraph PARALLEL1 ["runs in parallel"]
        UX["UI/UX · Sonnet 4.6\nFLOW → DESIGN → SPEC\nwireframes · component specs · all states"]
        SEC1["Security · Opus 4.6\nDESIGN-REVIEW\nauth model · threat surface · constraints"]
    end

    UX -->|specs| G2
    SEC1 -->|constraints| G2

    G2{{"Gate 2\ncleared to build"}}

    G2 -->|build| FS

    FS["Full Stack · Sonnet 4.6\nBUILD\ncode + auth + unit tests"]

    FS -->|working code| CP
    CP{{"Checkpoint\napp running locally · LGTM / FEEDBACK / STOP"}}

    CP -->|code to review| REV

    REV["Reviewer · Opus 4.6\nREVIEW\n7-specialist parallel army\nsecurity · performance · maintainability\nAPI · data · tests · errors"]

    REV -->|approved| G3

    G3{{"Gate 3\nreview verdict"}}

    G3 -->|code to test| QA
    G3 -->|code to audit| SEC2

    subgraph PARALLEL2 ["runs in parallel"]
        QA["QA · Sonnet 4.6\nTEST-RUN\ntest suite · auth matrix · edge cases"]
        SEC2["Security · Opus 4.6\nCODE-AUDIT\nOWASP · auth bypass · vulns"]
    end

    QA -->|test results| SHIPPER
    SEC2 -->|audit results| SHIPPER

    SHIPPER["Shipper · Sonnet 4.6\nSHIP\ntests → changelog → version bump → PR"]

    SHIPPER -->|ship report| RETRO

    RETRO["Retro · Sonnet 4.6\nRETRO\nwhat worked · what didn't · action items\nlearnings written to project memory"]

    RETRO -->|done| DONE

    DONE["Sprint Complete\nlearnings captured · ready for next sprint"]

The team

Navox Agents vs gstack

gstack (108K stars) is a great project — 23 workflow skills with a Bun+Chromium runtime. Different philosophy, different trade-offs.

Where gstack wins: larger community, browser-based visual tools, faster single-skill execution.

Where Navox wins: multi-agent orchestration with handoff contracts, automated quality gates, autonomous operation via SDK, zero dependencies, and the reliability guarantees that come from agents validating each other's work before passing it downstream.

The core difference: gstack gives you 23 independent tools. Navox gives you a team that works together.

Handoff contracts

Every agent has a handoff contract — a defined set of required sections it must include in its output before passing work to the next agent. Agents self-validate against their contract before completing.

This means:

• The Architect must include API contracts, auth model, and build order — not just a prose summary
• Full Stack must include a file manifest and run instructions — not just "I built it"
• Security must reference specific file paths and severity levels — not general advice
• QA must include exact pass/fail counts and reproduction steps — not approximations

If an upstream agent omits a required section, the downstream agent flags it before starting work. No agent guesses at what it should have received.

Full contract details: docs/handoff-chain.md

Builder philosophy

Every agent is guided by three principles from ETHOS.md:

1. Do the Complete Thing — no half-done work, no skipped edge cases
2. Investigate Before Acting — understand what exists before changing it
3. Builder Sovereignty — AI recommends, humans decide. Always.

These aren't decorative. They're enforced in every agent's prompt and checked by the eval system.

You stay in control

1. Agents pause at every gate and wait for your approval
2. Nothing destructive runs without your explicit sign-off
3. You can redirect, reject, or stop at any point

Agents stop. They wait. You decide. Then they continue.

Full guide: docs/hitl.md

Project memory

After each /agency-run, the team writes down what it learned in .claude/project-memory.md. Next run, it reads this file first — so it won't repeat work or ask you to re-explain the stack.

Each agent also keeps its own memory in .claude/memory/[agent].md with the same structure.

SDK — autonomous orchestration

The sdk/ directory contains a Python orchestration engine that runs sprint chains autonomously via the Anthropic API. Every agent step is evaluated against an 8/10 quality threshold — outputs below the bar are automatically retried with feedback.

cd sdk && pip install -e .

# Dry run — validate the chain without API calls
navox run full "Build an invoicing app" --dry-run

# Run a full sprint autonomously
navox run full "Build an invoicing app"

# Quick sprint or hotfix
navox run quick "Add dark mode toggle"
navox run hotfix "Fix 403 errors after login"

# Check journal status (resumable — interrupted sprints resume automatically)
navox status

Features:

• Eval-gated retries — deterministic grading after each step (8/10 threshold), auto-retry with feedback
• Content-addressed journaling — same task resumes from where it left off
• Parallel execution — agents in the same group run concurrently with failure containment
• Stop reason handling — all 6 Anthropic API stop reasons handled (end_turn, max_tokens, refusal, etc.)

Quality assurance

Every agent scores 10/10 against a quality rubric covering: frontmatter, modes, handoff contracts, anti-hallucination rules, anti-sycophancy, error handling, structured output, scope boundaries, ethos reference, and memory integration.

Validation — 383 checks across all agents, contracts, and registry:

bash scripts/validate.sh    # Structural checks
navox validate              # Full validation (Python)
navox score                 # 10-point rubric scoring

SDK tests — 91 tests covering orchestrator, journal, eval, validators, and contracts:

cd sdk && python -m pytest tests/ -q

What this is not

• Not a platform. No dashboard, no login.
• Not a SaaS. No subscription, no usage limit.
• Not a walled garden. The source is open — fork it, customize the prompts, make it yours.
• Not storing your data. Everything runs locally through Claude Code.
• Not black-box autonomous. The SDK runs autonomously but every step is evaluated, journaled, and resumable.

More from Navox Labs

Docs · Install · See it work · Report Bug

Built by Nahrin at Navox Labs · MIT License