NodeBench AI

The local-first operating-memory and entity-context layer for agent-native businesses.

Live: nodebenchai.com · MCP: npx nodebench-mcp · Docs: /developers

---

Evolution

PHASE 0 ─── Genesis
"Can we build operating intelligence for founders?"
    │
    ├── Collaborative editor + early AI agents (Convex + Vite)
    ├── LLM-as-a-judge validation (50% → 100% pass rate)
    └── VDR-pattern dossiers, SEC disambiguation, news feeds
    ↓
PHASE 1 ─── Search Intelligence
"What if every company had a structured intelligence packet?"
    │
    ├── Multi-source search fusion (Linkup + Gemini + web)
    ├── 6 role lenses: Founder, Investor, Banker, CEO, Legal, Student
    ├── Entity extraction with Gemini Structured Output
    └── 4-layer grounding pipeline (claim verification → citation chain)
    ↓
PHASE 2 ─── Agent Harness
"Search alone isn't enough — agents need to orchestrate tool chains."
    │
    ├── LLM-planned execution: plan → parallel dispatch → synthesize
    ├── Multi-provider failover: Gemini → OpenAI → Anthropic
    ├── Monte Carlo 3-case financial simulation (bull/base/bear)
    ├── Context budgeting (DeerFlow pattern: per-tool char budget)
    └── Kilo Code-style auto model routing by task complexity
    ↓
PHASE 3 ─── MCP Distribution
"Meet users where they are — inside Claude Code, Cursor, Windsurf."
    │
    ├── 350+ MCP tools across 50 domains
    ├── Progressive discovery: starter (19 tools) → load_toolset → full (350)
    ├── Lazy-loading toolset registry (62K lines, only parse on demand)
    ├── CLI subcommands: discover, setup, workflow, quickref, call, demo
    └── WebSocket MCP Gateway with API key auth + rate limiting
    ↓
PHASE 4 ─── Founder Platform
"Founders don't just research — they need operating clarity."
    │
    ├── 5-surface cockpit: Ask, Decisions, Research, Docs, System
    ├── Decision Workbench with HCSN timing hierarchy
    ├── Founder Dashboard: 9 tabs, artifact packets, weekly reset
    ├── Research Hub with live daily brief + signal tracking
    └── Entity relationship graph + competitive landscape
    ↓
PHASE 5 ─── Credibility & Trust Layer
"Good companies are judged on people, trust, and hidden criteria."
    │
    ├── "Why This Team" — parallel Gemini credibility assessment
    │     Founder credibility · Trust signals · Vision magnitude
    │     Reinvention capacity · Hidden qualification requirements
    ├── Local context injection (CLAUDE.md + memory + git history)
    ├── Post-extraction entity verification (proper noun + phrase check)
    └── Confidence formula: 20-95 range based on web source quality
    ↓
PHASE 6 ─── Agent Workspace
"Agents need persistent memory — skills, rules, tasks, research."
    │
    ├── 6 MCP workspace tools: write, read, list, mkdir, research, tasks
    ├── Implementation packet system (draft → approved → executing → completed)
    ├── WorkspaceExplorer UI: Agents → Workspace → Documents → Schedule → Roadmap
    ├── Local-first storage (~/.nodebench/workspace/) with platform sync
    └── All views agent-aware: Agent Artifacts, Agent Schedule, Agent Roadmap
    ↓
PHASE 7 ─── Intelligence → Execution Loop
"NodeBench decides WHAT to build. Claude Code executes."
    │
    ├── Implementation packets: objective, scope, constraints, validation
    ├── Claude Code bridge: structured prompt → Messages API → parsed result
    ├── Execution API: POST /api/execute → background dispatch → status polling
    ├── Validation layer: diff check, test check, constraint check, cost tracking
    └── PWA auto-reload on deploy (controllerchange listener)
    ↓
PHASE 8 ─── What's Next
    │
    ├── KAIROS-style tick loop: always-on background monitoring
    ├── MCP proxy integration layer (zero-code-change interception)
    ├── Trajectory replay: save paths, replay with checkpoints, fallback on drift
    ├── Workflow compression: learn shortest cheapest valid path
    └── Team/enterprise: shared context, approval queues, cost governance

---

Quick Start

# Clone and install
git clone https://github.com/HomenShum/nodebench-ai.git
cd nodebench-ai && npm install

# Set up environment
cp .env.example .env.local
# Add GEMINI_API_KEY (required for search)

# Run
npm run dev          # Frontend (Vite)
npx nodebench-mcp   # MCP server (standalone)

MCP Install (Claude Code / Cursor / Windsurf)

claude mcp add nodebench -- npx -y nodebench-mcp@latest

---

Architecture

┌─────────────────────────────────────────────────────────┐
│                  NODEBENCH AI APP                       │
│  Ask · Decisions · Research · Docs · Dashboard          │
│  Search → Intelligence Packet → Action                  │
└────────────────────────┬────────────────────────────────┘
                         │
┌────────────────────────┴────────────────────────────────┐
│              AGENT HARNESS + SEARCH ROUTE               │
│  LLM classification → Plan → Parallel execution →      │
│  Synthesis → Why This Team → Monte Carlo → Validation   │
└────────────────────────┬────────────────────────────────┘
                         │
┌────────────────────────┴────────────────────────────────┐
│                  MCP TOOL LAYER                         │
│  350 tools · 50 domains · Progressive discovery         │
│  web_search · linkup · enrich_entity · simulate · etc.  │
└────────────────────────┬────────────────────────────────┘
                         │
┌────────────────────────┴────────────────────────────────┐
│              WORKSPACE + EXECUTION                      │
│  Skills · Rules · Tasks · Research · Implementation     │
│  Claude Code bridge · Packet lifecycle · Validation     │
└────────────────────────┬────────────────────────────────┘
                         │
┌────────────────────────┴────────────────────────────────┐
│              PERSISTENCE                                │
│  Local: ~/.nodebench/ (SQLite + workspace files)        │
│  Cloud: Convex (documents, files, missions, agents)     │
│  Sync: local-first, Convex when available               │
└─────────────────────────────────────────────────────────┘

---

Key Surfaces

Surface	URL	What
Ask	/?surface=ask	Search any company, get structured intelligence with credibility layer
Decisions	/?surface=memo	Decision Workbench — map variables, compare branches, surface next move
Research	/?surface=research	Live daily brief, signal tracking, forecast context
Docs	/?surface=editor	Agent Artifacts, Workspace, Schedule, Roadmap
Dashboard	/?surface=ask&view=founder-dashboard	9-tab founder operating cockpit

---

Design DNA

• Glass cards: border-white/[0.06] bg-white/[0.02]
• Warm terracotta accent: #d97757
• Typography: Manrope (UI) + JetBrains Mono (data)
• Background: --bg-primary: #151413
• Section headers: text-[11px] uppercase tracking-[0.2em]

---

Stack

Layer	Technology
Frontend	React 19 + Vite + Tailwind CSS + Recharts
Backend	Convex (real-time DB + serverless functions)
Server	Express + WebSocket MCP Gateway
Search	Gemini 3.1 Flash + Linkup + web_search
Models	Gemini (primary) → OpenAI GPT-5.4 (fallback) → Anthropic (bridge)
MCP	350 tools, lazy-loaded, progressive discovery
Storage	SQLite (local) + Convex (cloud) + filesystem workspace
Deploy	Vercel (frontend + serverless) + PWA

---

License

MIT

---

Built by Homen Shum · Banking + Data Engineering + Agentic AI