Command Center

Centralized hub exposing all D:\ai-sandbox\ projects as callable tools from any MCP-aware AI client (opencode, Claude Desktop, Cursor, etc.).

Vision

From a single chat, trigger functionality across every project in the sandbox without caring which project handles it:

• "Convert report.docx to markdown" → convert_document
• "Render the intro video" → render_video
• "Analyze this CSV" → analyze_csv
• "Download this YouTube transcript" → download_youtube_subtitles
• "Format this markdown file" → format_document

How It Works (non-technical)

There are three ways to use Command Center. Pick the one that fits what you're doing.

Dashboard (web browser)

You open a web page in your browser. You see cards for each tool — analyze_csv, convert_document, render_video, etc. You fill in a form, click Run. The browser sends a request to a server on your machine. That server finds the right script, runs it (might spawn Python, ffmpeg, or other programs), waits for the result, and sends it back to your browser. The browser displays the result on screen.

The browser never runs anything directly. It's just a display. The server does all the work.

AI client (opencode / Claude Desktop / Cursor)

You type in natural language: "analyze my CSV at D:\data\costs.csv". The AI client asks Command Center what tools exist (auto-discovery via MCP protocol), picks the right one, calls it, and shows you the result. This works through a pipe between the AI client and Command Center — no web browser needed.

Scripts (curl, PowerShell, Python)

You call curl localhost:3010/tools/call with a JSON body describing which tool to run and what arguments. The HTTP API (a separate server on port 3010) receives it, runs the tool, and sends back the result. No browser, no AI client — useful for automation.

All three paths end up running the same tool scripts. Only the way you trigger them is different.

Surfaces

Surface	Port	Who uses it	How it works
Dashboard	3000	You (in a web browser)	Click cards, fill forms, hit Run — results display on screen
MCP server	stdio	AI clients (opencode, Claude Desktop, Cursor)	AI discovers tools automatically, calls them from chat
HTTP API	3010	Scripts (curl, Python, automation)	Send JSON requests, get JSON responses
Dev orchestrator	3000+3010	npm run dev or double-click start-dev.bat	Starts Dashboard and HTTP API together

Architecture

D:\ai-sandbox\
├── command_center/          ← This folder (the hub)
│   ├── server/
│   │   ├── index.mjs          MCP server entry (stdio)
│   │   └── http.mjs           HTTP API + SSE streaming (port 3010)
│   ├── tools/
   │   │   ├── ping.mjs
   │   │   ├── analyze_csv.mjs
   │   │   ├── compose_from_script.mjs
   │   │   ├── convert_document.mjs
   │   │   ├── download_youtube_subtitles.mjs
   │   │   ├── format_document.mjs
   │   │   ├── render_video.mjs
   │   │   └── ask.mjs
│   ├── dashboard/             Next.js 15 web GUI (port 3000)
│   │   ├── app/
│   │   ├── components/
│   │   └── ...
│   ├── sub_agents.json        Sub-agent registry (router agent)
│   ├── REGISTRY.md            Project inventory + tool mapping
│   ├── PROGRESS.md            Session log, decisions, next steps
│   └── AGENTS.md              Agent context (conventions, decisions)
│
├── csv_analyzer/             ← consumed project
├── markdown-formatter/
│   └── personal/             ← consumed project (port 3001)
├── vid/remotion/             ← consumed project
└── youtube-subtitle-download-plus-format/
    └── ...                     ← consumed project (port 3002)

Rule: command_center/ only contains orchestration code. Project code stays in its own folder, untouched.

Available Tools

Tool	Purpose	Wraps
ping	Health check (version + tool count)	—
analyze_csv	Run csv_analyzer — totals, daily breakdown, cost tiers	D:\ai-sandbox\csv_analyzer\
convert_document	Convert md/pdf/docx/html/txt (optional LLM formatting)	D:\ai-sandbox\markdown-formatter\personal\ (port 3001)
format_document	AI-format a markdown file in place	D:\ai-sandbox\markdown-formatter\personal\ (port 3001)
render_video	Render a Remotion composition to MP4	D:\ai-sandbox\vid\remotion
compose_from_script	Render a full video from a .md production script	D:\ai-sandbox\vid\remotion
download_youtube_subtitles	Fetch YouTube transcript, optionally LLM-format	D:\ai-sandbox\youtube-subtitle-download-plus-format\ (port 3002)
ask	Natural-language → tool routing via sub-agent	All tools above

Quick Start

Dashboard (Web GUI)

cd D:\ai-sandbox\command_center\dashboard
npm install
npm run dev

Open http://127.0.0.1:3000.

Dev Orchestrator (recommended for development)

cd D:\ai-sandbox\command_center
npm install
npm run dev

Or double-click start-dev.bat. Starts both the HTTP API and Dashboard together with color-coded output.

MCP Server (AI clients)

cd D:\ai-sandbox\command_center
npm install
node server/index.mjs

From any MCP client, the tools above are exposed via stdio.

HTTP API (programmatic)

cd D:\ai-sandbox\command_center
npm run http

Endpoints:

• GET / — server info
• GET /tools — full tool list with schemas
• GET /status — health check
• POST /tools/call — call any tool by name
• POST /ask — SSE streaming of the ReAct loop

Project Structure

command_center/
├── server/
│   ├── index.mjs              MCP server registration + tool handlers map
│   └── http.mjs               Express HTTP API + SSE streaming
├── tools/
│   ├── *.mjs                  One wrapper per tool (ESM, imports project binary)
│   └── ...
├── dashboard/                 Next.js 15 web app (see dashboard/README.md)
│   ├── app/
│   ├── components/
│   ├── lib/
│   └── ...
├── sub_agents.json            Sub-agent registry (model, tools, system prompt)
├── REGISTRY.md                Full project inventory + tool mapping
├── PROGRESS.md                Session log, decisions, open threads
├── AGENTS.md                  Agent context (auto-injected)
├── start-dev.bat              Double-click launcher for dev orchestrator
└── .ai/memory/                Project memory / brain

Contributing a New Tool

1. Add a row to REGISTRY.md inventory
2. Create tools/<name>.mjs wrapper (ESM, follow existing convention)
3. Register in server/index.mjs MCP tool list
4. Add to dashboard/lib/tools.ts if dashboard UI needed
5. Test: call from opencode chat or via the dashboard
6. Note the change in PROGRESS.md

Configuration

Environment variables (used by ask sub-agent):

Var	Purpose	Default
LLM_API_KEY	LLM auth (NVIDIA / NagaAI / OpenRouter / OpenAI)	Required for ask + dashboard chat
LLM_BASE_URL	LLM endpoint	https://integrate.api.nvidia.com/v1
LLM_MODEL	Model name	—
YT_LLM_BASE_URL	YouTube formatter LLM base URL	https://openrouter.ai/api/v1
YT_LLM_API_KEY	YouTube formatter LLM key	—
YT_LLM_MODEL	YouTube formatter model	google/gemma-4-31b-it:free

Sub-agent model can also be overridden per-agent in sub_agents.json.

See Also

• dashboard/README.md — Dashboard-specific docs
• REGISTRY.md — Full project inventory + tool mapping
• PROGRESS.md — Session log, decisions, open threads