AI-Powered Multi-Agent Development Team -- 8 specialist agents with a shared index system. Describe your app. They architect, design, code, test, and deploy it.
A plugin that turns any AI assistant into a coordinated development team. You describe what you want. 8 agents plan, build, test, and deploy it -- sharing context through a lightweight index system that saves 60-80% on tokens.
No templates. No boilerplate. The agents coordinate through a shared file index, reuse components, enforce design quality, and remember everything across sessions.
Avoids two confirmed Claude Code cache bugs (#40524, #34629) that can silently inflate API costs. The plugin's index-first architecture sidesteps both by design — especially the --resume bug, where /resumeproject is significantly cheaper.
npm install -g 10x-development-team
# Configure all your AI clients
10x-mcp setup
# Add plugin to any project
10x-mcp install-plugin ~/my-app
# Health check
10x-mcp doctorcd my-project
npm install 10x-development-team
# Postinstall auto-creates:
# .claude/ (agents, skills, knowledge, hooks)
# .10x/ (project index)
# .mcp.json (Playwright + 10x MCP)
# .env (skeleton)
# public/models/ (3D assets)npx 10x-development-team setup # Auto-detect all clients
npx 10x-development-team setup --client cursor # Specific client
npx 10x-development-team setup --all # All 9 clientsnpx 10x-development-team install-plugin ~/my-app
# Then: /start| Client | Setup Command |
|---|---|
| Claude Desktop | 10x-mcp setup --client claude-desktop |
| Claude Code | 10x-mcp install-plugin . or 10x-mcp setup --client claude-code |
| Cursor | 10x-mcp setup --client cursor |
| Windsurf | 10x-mcp setup --client windsurf |
| Cline | 10x-mcp setup --client cline |
| VS Code + Copilot | 10x-mcp setup --client vscode |
| Continue.dev | 10x-mcp setup --client continue |
| OpenAI Codex | 10x-mcp setup --client codex |
| OpenCode | 10x-mcp setup --client opencode |
Runs 100% locally. No hosting, no API keys, no cloud services.
The secret sauce. Agents don't scan your filesystem. They read a lightweight index.
| File | What It Tracks |
|---|---|
project.json |
App name, scope, stack, vision, status |
file-index.json |
Every file: type, description, exports, dependencies |
tasks.json |
Task tracker with goals, assignments, status |
feature-map.json |
Feature-to-file mapping with data flow wiring |
dev-log.md |
Chronological log of all agent actions |
assets.json |
Downloaded 3D models, textures, HDRIs with licenses |
| Table | What It Remembers |
|---|---|
projects |
Every project you've ever built |
sessions |
Build history, files created, tasks completed |
memories |
Decisions, preferences, user feedback |
shared_patterns |
Reusable code across projects |
SQLite is optional -- the server works without it (graceful fallback to .10x/ files).
| Without Index | With Index |
|---|---|
| Agent scans filesystem (~80K tokens) | Agent reads index (~2K tokens) |
| Each agent re-explores the codebase | Each agent knows exactly what exists |
| Duplicate components created | Component registry prevents duplication |
| API mismatches between front/back | API shapes documented in file-index |
| Session breaks lose context | Index persists, agents resume instantly |
Two confirmed, open bugs in Claude Code can break prompt caching. This plugin's architecture avoids both.
Bug 1: Conversation History Invalidation (#40524)
The standalone Claude Code binary can invalidate the cache when conversations contain CC internal strings. This is a niche trigger — most normal usage is unaffected.
How this plugin avoids the trigger: CLAUDE.md and knowledge files contain only clean project instructions -- never CC internals, billing strings, or source code.
Reported workaround: Use npx @anthropic-ai/claude-code instead of the standalone binary (community-reported).
Bug 2: --resume Cache Miss (#34629)
Since v2.1.69, every --resume causes a full cache rebuild. Cost scales with conversation size (~$0.15 on 500K tokens, proportionally less on smaller conversations).
How this plugin avoids it: The /resumeproject command reads small .10x/ index files (~5KB total) in a fresh session -- no --resume needed. This is the plugin's most concrete cost benefit.
| Method | Cost (500K token conversation) |
|---|---|
--resume |
~$0.15 one-time cache rebuild |
/resumeproject |
~$0.001 (reading index files) |
| Agent | Role | Scope |
|---|---|---|
| Team Lead | Plans, delegates, tracks progress, verifies design quality. Never writes code. | All |
| Frontend Dev | Pages, components, UI. 35+ components in registry. Design-system enforced. | All |
| Backend Dev | APIs, database schemas, auth, business logic. | MVP + Production |
| UI Designer | Design system: semantic tokens, gradients, shadows, animations, dark mode. | Prototype+ |
| 3D Designer | 3D scenes, particle systems, scroll-driven animations, WebGL effects (R3F, Drei, GSAP). | All |
| QA Tester | Type checks, build verification, visual quality audit, accessibility scan. | MVP + Production |
| Deployer | CI/CD pipelines, Docker, hosting config, performance headers. | Production |
| Error Recovery | Diagnoses build failures, runtime errors, dependency conflicts. Max 3 attempts. | On demand |
The plugin now treats production work like an engineering system, not just a code generator.
- Spec before code -- production work starts with an architecture / technical spec.
- Quality gates -- spec, design, implementation, QA, release readiness, and post-deploy verification are explicit checkpoints.
- Test pyramid -- unit, integration, and E2E coverage are planned separately instead of lumped into one vague "tests" step.
- Contract safety -- frontend/backend API shapes are verified and tracked in the file index.
- Migration safety -- schema changes require forward-safety and rollback notes.
- Performance budgets -- bundle and runtime targets are defined before release.
- Release discipline -- pre-deploy checklist, smoke tests, rollback plan, and post-deploy verification are part of the workflow.
- Security review -- production specs include threat modeling for the highest-risk abuse cases.
- Agent handoffs -- every specialist passes structured context to the next one.
This closes the gap between "multi-agent codegen" and how strong product teams actually ship.
| Scope | Output | Tech | Agents |
|---|---|---|---|
| Simple | Static pages | HTML + CSS + JS | Frontend only |
| Prototype | Clickable demo with mock data | Vite + React + Tailwind | UI + Frontend + 3D |
| MVP | Working app with real data | Next.js + Prisma + Auth | UI + Front + Back + 3D |
| Production | Full app, tested, deployed | Next.js + Tests + Docker | All 8 agents |
A landing page gets HTML. A SaaS gets Next.js. The plugin decides -- you don't have to.
Inspired by Lovable, v0, and Bolt.new -- output is polished, not just functional.
- Semantic tokens only --
bg-primary,text-foreground. Neverbg-blue-600,text-white. - Every interactive element has states -- hover, focus-visible, active, disabled + transitions.
- Every data view has 3 states -- loading skeleton, empty state, error state.
- Dark mode automatic -- semantic tokens handle light/dark.
- Mobile-first responsive -- design for 320px, enhance up.
- Rich animations -- Framer Motion, GSAP ScrollTrigger, CSS transitions.
- 3D experiences -- React Three Fiber, Drei, particles, scroll-driven 3D, glassmorphism.
- Accessible -- WCAG AA contrast, focus indicators, semantic HTML, aria attributes.
| Category | Tools |
|---|---|
| Project Lifecycle | tenx_start, tenx_build, tenx_projects, tenx_read_index, tenx_update_index |
| Knowledge | tenx_get_skill, tenx_get_knowledge, tenx_list_knowledge, tenx_get_components |
| Memory & Context | tenx_save_memory, tenx_get_context, tenx_get_agent |
| Session & Health | tenx_update_session, tenx_health |
| User Interaction | tenx_ask_user, tenx_save_answer |
| 3D Asset Pipeline | tenx_download_asset, tenx_list_assets, tenx_browse_3d_sources, tenx_model_sources |
| Storyboard & Visual | tenx_storyboard, tenx_screenshot |
| Command | Description |
|---|---|
/start |
New project -- describe what you want, agents build it |
/build |
Execute the full build pipeline |
/add-page |
Add a page to your app |
/add-feature |
Add a feature (frontend + backend) |
/connect-data |
Wire external data (API, Supabase, Firebase, Sheets, CSV) |
/modify-ui |
Quick UI changes -- theme, layout, dark mode, animations |
/generate |
Generate component, API route, hook, model, or test |
/fix |
Describe the bug, plugin locates and fixes it |
/refactor |
Improve code quality without changing behavior |
/review |
Code quality + visual quality review |
/explain |
Explain how any part of the codebase works |
/deploy |
Set up deployment (Vercel, Docker, CI/CD) |
/resumeproject |
Continue where you left off -- avoids --resume cache bug (#34629) |
/projects |
List, switch, manage all your projects |
/status |
Project dashboard |
/update-deps |
Check and update dependencies safely |
/config |
View or change project settings |
/index |
View or rebuild the file index |
/help |
Show all commands |
In Claude Code, type /start to begin. All commands use short names: /build, /add-page, /fix, etc.
Agents don't invent code. They copy proven patterns from the knowledge base:
| Category | Files | Examples |
|---|---|---|
| Frameworks | 5 | Next.js, Vite+React, HTML/CSS/JS, Expo |
| Libraries | 15 | shadcn/ui, Tailwind, Framer Motion, Prisma, Stripe, R3F, Drei, GSAP |
| Patterns | 23 | Design system, 3D scenes, particles, scroll-driven 3D, glassmorphism, auth, SEO |
| Components | 18 | Button, card, navbar, sidebar, modal, data-table, auth pages, 3D scene, particles |
| Source | What You Get | License |
|---|---|---|
| Poly Haven | Models, HDRIs, textures | CC0 |
| Kenney | Low-poly game assets | CC0 |
| Sketchfab | High-quality models | CC-BY/CC0 |
| Mixamo | Animated characters | Free |
| Quaternius | Cartoon characters/nature | CC0 |
Use tenx_browse_3d_sources to search across all sources. Use tenx_download_asset to add to your project.
For all clients except VS Code:
{
"mcpServers": {
"10x-dev": {
"command": "npx",
"args": ["-y", "10x-development-team"]
}
}
}For VS Code + GitHub Copilot (.vscode/mcp.json):
{
"servers": {
"10x-dev": {
"command": "npx",
"args": ["-y", "10x-development-team"]
}
}
}User (any AI client)
|
+-- Claude Desktop -----> MCP Server (stdio, local)
+-- Claude Code --------> Direct plugin (.claude/) OR MCP Server
+-- Cursor IDE ---------> MCP Server (stdio, local)
+-- Windsurf IDE -------> MCP Server (stdio, local)
+-- Cline (VS Code) ---> MCP Server (stdio, local)
+-- VS Code + Copilot --> MCP Server (stdio, local)
+-- Continue.dev -------> MCP Server (stdio, local)
+-- Codex CLI ----------> MCP Server (stdio, local)
+-- OpenCode -----------> MCP Server (stdio, local)
|
+-- All modes ----------> ~/.10x/memory.db (SQLite -- global cross-project memory)
+-- All modes ----------> .10x/ per project (index, tasks, features, dev log)
| Model Size | Behavior |
|---|---|
| Haiku / Sonnet | COPY code from knowledge files. Change only names, props, content. |
| Opus | USE knowledge as base. ENHANCE with animations, a11y, dark mode, SEO. |
If you are an AI agent that has been asked to set up or use the 10x Development Team:
- Install:
npx 10x-development-team setup --client [your-client] - The MCP server exposes 22 tools. Call
tenx_startto begin a new project. - Call
tenx_read_indexto understand current project state before making changes. - Call
tenx_get_knowledgewithcategoryandfile_namebefore writing code. - Call
tenx_get_agentwithagent_nameto load specialist agent instructions. - Call
tenx_update_indexafter creating/modifying files to keep the index in sync. - Call
tenx_healthto verify setup is working correctly. - The system works 100% locally -- no API keys, no hosting, no cloud services.
- Homepage: 10x-anit.github.io/10x-development-team
- GitHub: github.com/10x-Anit/10x-development-team
- npm: npmjs.com/package/10x-development-team
MIT License | Built by 10x-Anit