Nodebench

NodeBench MCP

[npm version] [npm downloads] [License: MIT] [GitHub stars] [MCP Compatible] [Default]

Investigate a topic and return a sourced report fast. The NodeBench MCP architecture is now split into three install lanes that share one runtime.

Default install: 9 visible tools total, including 7 core workflow tools: investigate, compare, track, summarize, search, report, and ask_context.

# Core workflow laneclaude mcp add nodebench -- npx -y nodebench-mcp
# Power laneclaude mcp add nodebench-power -- npx -y nodebench-mcp-power
# Admin laneclaude mcp add nodebench-admin -- npx -y nodebench-mcp-admin

What's New

• v3 default surface - the default preset is now a workflow-first facade instead of a tool warehouse.
• Admin-only runtime gates - dashboards and the observability watchdog no longer start on the default hot path. Use --admin, --dashboards, or --watchdog explicitly.
• Faster health path - --health now loads only the active preset instead of eagerly loading every domain.
• Separate install lanes - nodebench-mcp-power and nodebench-mcp-admin are now published companion packages, while --preset power and --preset admin remain compatible paths on the core package.

New here? Read AGENT_LOGIC.md for the complete guide to how NodeBench thinks.

RETHINK REDESIGN APR 2026

This section records a repo-grounded reexamination of NodeBench MCP as of April 2026. It is intentionally blunt. The goal is to shrink the product to clear, measurable workflows and make the runtime and claims trustworthy.

Repo-grounded findings

• Surface area and messaging were out of sync
  • At the time of this audit, the README marketed 350+ tools, founder under 50 tools, and full as 338.
  • The audit that drove this section measured roughly 28 tools in the starter tools/list payload, roughly 186 in founder, and roughly 546 in full, before counting the extra dynamic-loading helpers separately.
• The core executable is overloaded
  • packages/mcp-local/src/index.ts currently combines MCP serving, analytics tracking, embedding bootstrapping, profiling hooks, A/B instrumentation, dynamic tool loading, dashboard startup, and engine hosting in one runtime path.
• Performance is not yet credibly measured
  • During this audit, node packages/mcp-local/dist/index.js --health took about 2.9s locally because CLI mode eagerly loaded all toolsets before exit.
  • The main performance comparison in the package measures local toolchain overhead, not end-user outcome quality, startup SLOs, workflow completion, or real provider spend.
• Cost reporting is not yet auditable
  • The current profiler relies on heuristic per-tool cost estimates, not authoritative provider billing or exact token accounting.
• Workflow clarity is weak
  • The product currently explains presets, domains, discovery engines, and meta-systems before it proves one concrete job to be done.

Redesign principles

1. One dominant workflow per persona
   • A founder should land on one obvious path.
   • A banker should land on one obvious path.
   • The MCP should prove value before it explains its architecture.
2. Truthful counts and truthful claims
   • Preset docs, README counts, and tools/list output must agree.
3. Speed is a product behavior
   • Measure startup time, tools/list payload size, per-tool latency, and workflow time-to-first-value.
4. Keep optional systems out of the hot path
   • Embeddings, dashboards, profilers, and engine surfaces should not make the default MCP startup slower.
5. Prefer workflows over catalogs
   • Opinionated, repeatable outcomes are more valuable than an oversized flat tool inventory.
6. Ship process, not prose
   • The right mental model is a small number of strong workflows with evidence and verification, not a giant capabilities brochure.

Execution board

Reference models

• agent-skills: process, not prose. Strong lifecycle commands, checkpoints, and evidence requirements.
• GitHub MCP Server: narrower toolsets improve tool choice and reduce context size. The server also exposes a simple tool-search utility.
• Google MCP Toolbox: performance and operability are anchored in concrete runtime primitives such as connection pooling and OpenTelemetry, not only internal heuristics.

Bottom line

NodeBench MCP should evolve from a monolithic "hundreds of tools" surface into a smaller number of opinionated workflow products backed by a measurable, trustworthy runtime.

Unified cross-surface spec: docs/architecture/UNIFIED_WORKFLOW_SPEC.md

Production companion spec: docs/architecture/UNIFIED_WEB_MCP_PRODUCTION_SPEC.md

Founder MCP flow spec: docs/architecture/FOUNDER_MCP_FLOW_SPEC.md

What You Get

NodeBench is now a workflow-first MCP. The default install proves one concrete job quickly, and the heavier surfaces are still available when you explicitly ask for them.

Install Lanes And Presets

Default workflow

ASK -> CHECK -> WRITE -> SAVE

The default preset is optimized for that loop. It does not start local dashboards or the observability watchdog unless you pass admin flags explicitly.

# Claude Code core laneclaude mcp add nodebench -- npx -y nodebench-mcp
# Claude Code power laneclaude mcp add nodebench-power -- npx -y nodebench-mcp-power
# Claude Code admin laneclaude mcp add nodebench-admin -- npx -y nodebench-mcp-admin
# Windsurf / Cursor - add --preset to args only when you want a compatibility preset on the core package

Quick Start

Claude Code (CLI)

claude mcp add nodebench -- npx -y nodebench-mcpclaude mcp add nodebench-power -- npx -y nodebench-mcp-powerclaude mcp add nodebench-admin -- npx -y nodebench-mcp-admin

Or add to ~/.claude/settings.json or .mcp.json in your project root:

{  "mcpServers": {    "nodebench": {      "command": "npx",      "args": ["-y", "nodebench-mcp"]    }  }}

Cursor

Add to .cursor/mcp.json (or Settings > MCP). Use the cursor preset to stay within Cursor's tool limit:

{  "mcpServers": {    "nodebench": {      "command": "npx",      "args": ["-y", "nodebench-mcp", "--preset", "cursor"]    }  }}

Windsurf

Add to .windsurf/mcp.json (or Settings > MCP > View raw config):

{  "mcpServers": {    "nodebench-power": {      "command": "npx",      "args": ["-y", "nodebench-mcp-power"]    }  }}

Other MCP Clients

Any MCP-compatible client works. Point command to npx and choose the package that matches the lane you want:

• ["-y", "nodebench-mcp"] for the tiny default workflow lane
• ["-y", "nodebench-mcp-power"] for the expanded founder/research lane
• ["-y", "nodebench-mcp-admin"] for the operator lane

Add "--preset", "<name>" only when you want a compatibility preset on the core package.

First Prompts to Try

# Research a topic> Use investigate with topic="Anthropic" to produce a sourced report
# Compare two entities> Use compare with entities=["Anthropic","OpenAI"] to get a side-by-side brief
# Turn rough notes into a report> Use report with topic="AI agent infrastructure" and context="..." to produce a decision memo
# Search live web + saved knowledge> Use search with query="MCP server best practices 2026"
# Track an entity> Use track with action="add" and entity="Anthropic"
# Expand only when you need more> Use discover_tools("visual QA for a Vite app"), then load_toolset("ui_ux_dive")

Optional: API Keys

export GEMINI_API_KEY="your-key"        # Web search + vision (recommended)export GITHUB_TOKEN="your-token"        # GitHub (higher rate limits)

Set these as environment variables, or add them to the env block in your MCP config:

{  "mcpServers": {    "nodebench": {      "command": "npx",      "args": ["-y", "nodebench-mcp"],      "env": {        "GEMINI_API_KEY": "your-key",        "GITHUB_TOKEN": "your-token"      }    }  }}

Progressive Discovery — How Optional Toolsets Stay Off the Hot Path

The default preset exposes 9 tools. Everything else stays off the hot path until you deliberately load a specific toolset.

How it works

1. discover_tools("your task description")    → ranked results from the full registry2. load_toolset("ui_ux_dive")                 → a specific toolset activates in your session3. Use the newly loaded tools directly        → no proxy, native binding4. Keep the default surface small             → only load what the workflow needs

Multi-modal search engine

discover_tools scores tools using 14 parallel strategies:

Plus cursor pagination (offset/limit), result expansion (expand: N), and multi-hop BFS traversal (depth: 1-3) via get_tool_quick_ref.

Client compatibility

Key Features

Decision Intelligence (Deep Sim)

Simulate decisions before committing. Run scenarios with multiple variables, score trajectories, generate postmortems, produce decision memos.

Causal Memory

Track actions, paths, and state across sessions. Important-change review surfaces what shifted since your last session.

Artifact Packets

Every analysis produces a shareable artifact — decision memos, delegation briefs, investigation reports. The output is the distribution.

Founder Tools

Weekly reset, pre-delegation briefing, company tracking, important-change review. Built for the founder who needs to make 20 decisions a day with incomplete information.

Knowledge Compounding

record_learning + search_all_knowledge — findings persist across sessions. By session 9, the agent finds 2+ relevant prior findings before writing a single line of code.

Headless Engine API

NodeBench ships a headless, API-first engine for programmatic access.

# Start MCP server with engine API on port 6276npx nodebench-mcp --engine
# With auth tokennpx nodebench-mcp --engine --engine-secret "your-token"

Fine-Grained Control

# Include only specific toolsetsnpx nodebench-mcp --toolsets deep_sim,recon,learning
# Exclude heavy toolsetsnpx nodebench-mcp --exclude vision,ui_capture,parallel
# Dynamic loading — start minimal, load on demandnpx nodebench-mcp --dynamic
# Smart preset recommendation based on your projectnpx nodebench-mcp --smart-preset
# Usage statsnpx nodebench-mcp --stats
# List all presetsnpx nodebench-mcp --list-presets
# See all optionsnpx nodebench-mcp --help

TOON Format — Token Savings

TOON (Token-Oriented Object Notation) is on by default. Every tool response is TOON-encoded for ~40% fewer tokens vs JSON. Disable with --no-toon.

Security & Trust Boundaries

NodeBench MCP runs locally on your machine.

• Local runtime data lives under ~/.nodebench/. Native SQLite is used when the addon is available; lightweight mode skips the native persistence layer. No data is sent to external servers unless you provide API keys and use tools that call external APIs.
• Analytics data never leaves your machine.
• The local_file toolset can read files anywhere your Node.js process has permission. Use the default preset to keep local-file tools off the hot path.
• All API keys read from environment variables — never hardcoded or logged.
• All database queries use parameterized statements.

Build from Source

git clone https://github.com/HomenShum/nodebench-ai.gitcd nodebench-ai/packages/mcp-localnpm install && npm run build

Then use absolute path:

{  "mcpServers": {    "nodebench": {      "command": "node",      "args": ["/path/to/packages/mcp-local/dist/index.js"]    }  }}

Troubleshooting

"No search provider available" — Set GEMINI_API_KEY, OPENAI_API_KEY, or PERPLEXITY_API_KEY

"GitHub API error 403" — Set GITHUB_TOKEN for higher rate limits

"Cannot find module" — Run npm run build in the mcp-local directory

MCP not connecting — Check path is absolute, run claude --mcp-debug, ensure Node.js >= 18

Windsurf not finding tools — Verify ~/.codeium/windsurf/mcp_config.json has correct JSON structure

Cursor tools not loading — Ensure .cursor/mcp.json exists in project root. Use --preset cursor to stay within the tool cap. Restart Cursor after config changes.

Dynamic loading not working — Claude Code and GitHub Copilot support native dynamic loading. For Windsurf/Cursor, prefer --preset cursor or --preset power if your client does not refresh tools reliably after load_toolset.

License

MIT