Command Center

A command space for your agent. You review exceptions, it does the rest.

Command Center is an opinionated workspace your AI agent drives on a schedule. Your agent captures, clusters, triages, promotes, and archives items through a defined state machine. You get a dashboard for inspecting exceptions, duplicates, and progress — not a manual inbox to clean.

It ships with a Next.js dashboard, a SQLite-backed store, a REST API, and an MCP server so agents can operate natively. A reference openclaw setup is included for cron-driven daily triage and weekly consolidation.

This is a personal project I maintain on hobby time. Fork it, steal ideas, open issues — but there's no SLA. PRs welcome.

⚠ Security — read before deploying

Command Center has no authentication by design. It assumes you are the only caller. The default dev/start scripts bind to 127.0.0.1 so nothing outside your machine can reach the API.

Do not expose this to a public network. Running it on a public IP, or forwarding the port without auth in front, lets anyone on the internet read, modify, delete, and promote your captures — and trigger outbound webhook calls from your server.

For remote access, tunnel over SSH (ssh -L 3005:localhost:3005 …) or put it behind a reverse proxy that enforces auth (Caddy + basicauth, Tailscale, Cloudflare Access, etc.). Don't put captures you'd be uncomfortable sharing into the workspace — the promoted item's full content is sent to the webhook target.

What it looks like

• Dashboard (http://localhost:3005): exception-first view of your workspace — today's priorities, blockers, stale items, lane status, pending decisions, waiting-on follow-ups, 24h pulse.
• State machine: every item flows raw → clustered → candidate → promoted, with reference and archived as terminal states.
• Lanes: items are sorted into configurable strategic lanes (Work / Personal / Later by default, or whatever you define).
• MCP tools: get_workspace, update_items, promote_items, scan_queue_noise, and six more — all thin wrappers over the REST API so agents can read and write the whole workspace in one trip.

Quickstart

git clone https://github.com/jendrypto/command-center.git
cd command-center

npm install
cp .env.example .env

# Edit command-space.config.ts to set your agent's name, lanes, and promotion target.
# Edit .env to set your webhook URL (if using the default webhook promotion).

npm run dev

Open http://localhost:3005. You'll see an empty dashboard.

Wire your agent (openclaw)

One-time setup for openclaw users — see examples/openclaw/README.md for the full walkthrough. Short version:

# Build the MCP server
npm run mcp:build

# Register it with openclaw
./examples/openclaw/install-mcp.sh

# Paste the three openclaw bootstrap snippets into your workspace
# (SOUL.md, AGENTS.md, TOOLS.md)

# Install the daily triage cron
./examples/openclaw/install-daily-triage.sh
./examples/openclaw/install-weekly-consolidation.sh

Your agent will run a triage pass every weekday morning. You open the dashboard to see what it did and what it kicked up for review.

Wire a different agent

The MCP server speaks standard Model Context Protocol over stdio, so any MCP-aware client (Claude Desktop, Claude Code, Cursor, Zed, openclaw, etc.) can use it. For non-MCP agents, the REST API at /api/agent is a straight JSON contract — see HOWTO.md.

Configuration

Edit command-space.config.ts. That file is the only per-install surface you control:

{
  agent: {
    name: "Scout",                // Shown in the dashboard
    purpose: "...",               // One-line description
  },
  focusAreas: [
    { id: "work", label: "Work", color: "#6c2bee", priority: 100 },
    // ...
  ],
  promotionTarget: {
    type: "webhook",              // or "none"
    url: process.env.PROMOTION_WEBHOOK_URL,
    headers: { Authorization: `Bearer ${process.env.PROMOTION_WEBHOOK_TOKEN}` },
  },
  dashboard: { title: "Command Center" },
}

States, dispositions, metadata fields, and the dashboard layout are product-level decisions and live in the code. If you need to change those, fork the repo. Everything else is config.

State machine

   ┌─────┐   cluster    ┌───────────┐   surface    ┌───────────┐   promote   ┌──────────┐
──▶│ raw │─────────────▶│ clustered │─────────────▶│ candidate │────────────▶│ promoted │
   └──┬──┘              └─────┬─────┘              └─────┬─────┘             └──────────┘
      │                       │                          │
      │                       │                          └───▶ reference
      └──────────▶ archived ◀─┴──▶ archived

• raw: new capture, no classification yet
• clustered: tied to a theme via cluster_key
• candidate: needs human judgment or agent confidence is low
• promoted: pushed to an external system (via webhook or your chosen adapter)
• reference: kept but not action-driving
• archived: dead, merged duplicate, or resolved

Agent metadata

The agent writes these fields on every triage pass. Your dashboard surfaces them.

Field	Purpose
reviewed_at	When the agent last touched the item
agent_confidence	0–1 confidence in the disposition
disposition	keep_incubating, connect_cluster, promote, reference, archive, merge_duplicate
duplicate_of	If this item is a duplicate, the canonical item's id
cluster_key	Free-form string grouping related items
promotion_target	Optional target label passed to the promotion adapter
needs_review	Boolean — set true when human judgment is needed
attention_reason	Short sentence explaining why it needs review
focus_area	Which lane this belongs to (from config)
focus_score	Priority score (derived from lane by default)

API contract

For full MCP/REST contracts see HOWTO.md. Quick reference:

• GET /api/agent — workspace snapshot (stats, attention queue, duplicates, backlog, connections)
• POST /api/agent — batch updates/connections/disconnects/promotions in one call
• GET /api/queue-cleaner — noise report (duplicates, stale groups, heartbeat clutter, orphans)
• POST /api/queue-cleaner — cleanup actions
• POST /api/promote — promote a single item to the configured target

Repo layout

command-center/
├── app/                 # Next.js dashboard + REST API
├── components/          # React UI
├── lib/                 # Core logic (db, config, promotion, queue-cleaner)
├── mcp-server/          # Stdio MCP server wrapping the REST API
├── examples/openclaw/   # Drop-in bootstrap snippets + cron scripts for openclaw
├── command-space.config.ts   # Per-install config (edit me)
├── .env.example
├── HOWTO.md             # Extending: adapters, agents, API
└── README.md            # You are here

Extending

See HOWTO.md:

• Writing a custom promotion adapter (Notion, Linear, your own system)
• Using a non-openclaw agent (raw REST contract)
• Adding capture sources (browser, CLI, email-in)
• Customizing the dashboard

License

MIT — see LICENSE.